predictionet (version 1.18.0)

netinf: Function performing network inference by combining priors and genomic data

Description

Main function of the predictionet package, netinf infers a gene network by combining priors and genomic data. The two main network inference methodologies implemented so far are the bayesian and regression-based inferences.

Usage

netinf(data, categories, perturbations, priors, predn, priors.count = TRUE, priors.weight = 0.5, maxparents = 3, subset, method = c("regrnet", "bayesnet"), ensemble = FALSE, ensemble.model = c("full","best"), ensemble.maxnsol = 3, causal=TRUE, seed, bayesnet.maxcomplexity=0, bayesnet.maxiter=100, verbose = FALSE)

Arguments

data

matrix of continuous or categorical values (gene expressions for example); observations in rows, features in columns.

categories

if this parameter missing, 'data' should be already discretized; otherwise either a single integer or a vector of integers specifying the number of categories used to discretize each variable (data are then discretized using equal-frequency bins) or a list of cutoffs to use to discretize each of the variables in 'data' matrix. If method='bayesnet' and categories is missing, data should contain categorical values and the number of categories will determine from the data.

perturbations

matrix of 0,1 specifying whether a gene has been perturbed (e.g., knockdown, overexpression) in some experiments. Dimensions should be the same than data.

priors

matrix of prior information available for gene-gene interaction (parents in rows, children in columns). Values may be probabilities or any other weights (citations count for instance). if priors counts are used the parameter priors.count should be TRUE so the priors are scaled accordingly.

predn

indices or names of variables to fit during network inference. If missing, all the variables will be used for network inference. Note that for bayesian network inference (method='bayesnet') this parameter is ignored and a network will be generated using all the variables.

priors.count

TRUE if priors specified by the user are number of citations (count) for each interaction, FALSE if probabilities or any other weight in [0,1] are reported instead.

priors.weight

real value in [0,1] specifying the weight to put on the priors (0=only the data are used, 1=only the priors are used to infer the topology of the network).

maxparents

maximum number of parents allowed for each gene.

subset

vector of indices to select only subset of the observations.

method

regrnet for regression-based network inference, bayesnet for bayesian network inference with the catnet package.

ensemble

TRUE if the ensemble approach should be used, FALSE otherwise.

ensemble.model

Could be either full or best depending how the equivalent networks are selected to be included in the ensemble network: for full bootstrapping is used to identify all the statistically equivalent networks, it best only the top ensemble.maxnsol are considered at each step of the feature selection.

ensemble.maxnsol

maximum number of solutions to consider at each step of the feature selection for the method=ensemble.regrnet, default is 3.

causal

'TRUE' if the causality should be inferred from the data, 'FALSE' otherwise

seed

set the seed to make the network inference deterministic.

bayesnet.maxcomplexity

maximum complexity for bayesian network inference, see Details.

bayesnet.maxiter

maximum number of iterations for bayesian network inference, see Details.

verbose

TRUE if messages should be printed, FALSE otherwise.

Value

method

name of the method used for network inference.

ensemble

is the network build using the ensemble approach?

topology

adjacency matrix representing the topology of the inferred network; parents in rows, children in columns.

topology.coeff

if method='regrnet' topology.coeff contains an adjacency matrix with the coefficients used in the local regression model; parents in rows, children in columns. Additionally the beta_0 values for each model in the first row of the matrix

edge.relevance

relevance score for each edge (see Details).

Details

bayesnet.maxcomplexity and bayesnet.maxiter are parameters to be passed to the network inference method (see cnSearchOrder and cnSearchSA from the catnet package for more details).

Relevance score is either MRMR scores if causal=FALSE or causality score if causal=FALSE.

Examples

## load gene expression data for colon cancer data, list of genes related to RAS signaling pathway and the corresponding priors
data(expO.colon.ras)
## create matrix of perturbations (no perturbations in this dataset)
pert <- matrix(0, nrow=nrow(data.ras), ncol=ncol(data.ras), dimnames=dimnames(data.ras))

## number of genes to select for the analysis
genen <- 10
## select only the top genes
goi <- dimnames(annot.ras)[[1]][order(abs(log2(annot.ras[ ,"fold.change"])), decreasing=TRUE)[1:genen]]
mydata <- data.ras[ , goi, drop=FALSE]
myannot <- annot.ras[goi, , drop=FALSE]
mypriors <- priors.ras[goi, goi, drop=FALSE]
mydemo <- demo.ras
mypert <- pert[ , goi, drop=FALSE]

########################
## regression-based network inference
########################
## infer global network from data and priors
mynet <- netinf(data=mydata, perturbations=mypert, priors=mypriors, priors.count=TRUE, priors.weight=0.5, maxparents=3, method="regrnet", seed=54321)

## plot network topology
mytopo <- mynet$topology
library(network)
xnet <- network(x=mytopo, matrix.type="adjacency", directed=TRUE, loops=FALSE, vertex.attrnames=dimnames(mytopo)[[1]])
plot.network(x=xnet, displayisolates=TRUE, displaylabels=TRUE, boxed.labels=FALSE, label.pos=0, arrowhead.cex=2, vertex.cex=4, vertex.col="royalblue", jitter=FALSE, pad=0.5)

## export network as a 'gml' file that you can import into Cytoscape
## Not run: rr <- netinf2gml(object=mynet, file="/predictionet_regrnet")

########################
## bayesian network inference
########################
## discretize gene expression values in three categories
categories <- rep(3, ncol(mydata))
## estimate the cutoffs (tertiles) for each gene
cuts.discr <- lapply(apply(rbind("nbcat"=categories, mydata), 2, function(x) { y <- x[1]; x <- x[-1]; return(list(quantile(x=x, probs=seq(0, 1, length.out=y+1), na.rm=TRUE)[-c(1, y+1)])) }), function(x) { return(x[[1]]) })
mydata <- data.discretize(data=mydata, cuts=cuts.discr)

## infer a bayesian network network from data and priors
## Not run: mynet <- netinf(data=mydata, perturbations=mypert, priors=mypriors, priors.count=TRUE, priors.weight=0.5, maxparents=3, method="bayesnet", seed=54321)

## plot network topology
## Not run: mytopo <- mynet$topology
## Not run: library(network)
## Not run: xnet <- network(x=mytopo, matrix.type="adjacency", directed=TRUE, loops=FALSE, vertex.attrnames=dimnames(mytopo)[[1]])
## Not run: plot.network(x=xnet, displayisolates=TRUE, displaylabels=TRUE, boxed.labels=FALSE, label.pos=0, arrowhead.cex=2, vertex.cex=4, vertex.col="royalblue", jitter=FALSE, pad=0.5)

## export network as a 'gml' file that you can import into Cytoscape
## Not run: rr <- netinf2gml(object=mynet, file="/predictionet_bayesnet")