create_signetwork: inhibMCMC Analysis: find significant edges from independent...

In ddepn: Dynamic Deterministic Effects Propagation Networks: Infer signalling networks for timecourse RPPA data.

Description

Uses statistical tests (Wilcox Rank-Sum or T-test) to find significant edges from a number of independent MCMC runs (performed by passing inference="mcmc" to the ddepn function call).

Usage

create_signetwork(ret, alpha = 0.05, adj.method = "BY", plot = FALSE, 
				type = "wilcox", alternative = "one.sided", 
				paired = FALSE, ord = NULL)

create_signetwork_cv(ret, alpha = 0.05, adj.method = "BY", plot = FALSE, 
				type = "wilcox", alternative = "one.sided", 
				paired = FALSE, ord = NULL, sel_policy = "strict")

perform_network_tests(ret, alpha=0.05, plot=FALSE, type="wilcox", 
					alternative="one.sided", paired=FALSE, ord=NULL)

perform_edge_test(ret, from, to, alpha=0.05, plot=TRUE, type="wilcox", 
					alternative="one.sided", paired=FALSE)

draw_segments(tx, rng, alpha)

Arguments

ret	List. Either the returned object of the ddepn call using the inhibMCMC inference (argument inference="mcmc"), or the sublist samplings from this object.
alpha	Numeric in [0;1]. Significance level for the tests.
adj.method	String. Adjustment for multiple testing. See function p.adjust from the stats package for p-value adjustment methods. Most common settings are "none", "BH" or "none".
plot	Boolean. Should boxplots of the activation/inhibition edge counts (respectively confidences) be plotted?
type	String. Which test to use, either "wilcox" or "ttest".
alternative	String. Alternative to test. One out of "one.sided", "two.sided".
paired	Boolean. Paired test? Should normally be unpaired, i.e. FALSE.
ord	Vector of strings. Defines a node ordering.
from	String. Defines the source node, from which an edge originates. Used in perform_edge_test and must be one of the node names of the components in the network.
to	String. Defines the target node, to which an edge points. Used in perform_edge_test and must be one of the node names of the components in the network.
tx	Numeric Vector. Holds the pvalue for alternative greater and alternative less.
rng	The range of the values that were tested against each other.
sel_policy	String. If "strict", an edge is included in the final network if it occurred in all CV runs. If "medium", an edge is included if it occurs in more than half of the CV runs. If "lenient", an edge is included if it occurs in any of the CV runs.

Details

The main function is create_signetwork.
Using multiple inhibMCMC chains (a number L), a significance testing procedure is used to calculate signfificantly occurring edges. For each edge, the number of sampled activations and inhibitions is counted in each run, divided by the total number of sampled edges, resulting in a 'confidence' for each edge for each run. To find significant edges, a statistical test is performed (usually Wilcoxon Rank Sum test) to test the NULL that the means of the L activation confidences (c_a) equal the mean of the L inhibition confidences (c_i). The alternatives c_a > c_i and c_a < c_i are used to determine the type of interaction.

In create_signetwork_cv a leave-one-out crossvalidation is performed to test for significant edges, i.e. every MCMC chain is left out once and the create_signetwork is called for the remaining runs. A final network is assembled from the CV networks according to the edge selection policy given in argument sel_policy. Note that at least 4 independent inhibMCMC chains have to be present to use this function.

The functions perform_network_tests, perform_edge_test and draw_segments are usually not called directly.

Value

An adjacency matrix containing the significant edges.

Note

Note that create_signetwork can only be performed for the results of the MCMC inference types and also only if multiple chains were run.

Author(s)

Christian Bender

References

Bender 2011, Systematic analysis of time resolved high-throughput data using stochastic network inference methods, Dissertation, University of Heidelberg, Faculty of Biology

See Also

ddepn, mcmc_ddepn

Examples

## Not run: 
## load package
library(ddepn)
library(multicore)

## sample a network and data
set.seed(1234)
n <- 6
signet <- signalnetwork(n=n, nstim=2, cstim=0, prop.inh=0.2)
phit <- signet$phi
stimuli <- signet$stimuli
dataset <- makedata(phit, stimuli, mu.bg=1200, sd.bg=400,
				mu.signal.a=2000, sd.signal.a=1000)

## use original network as prior matrix
## reset all entries for inhibiting edges 
## to -1
B <- phit
B[B==2] <- -1

## perform inhibMCMC inference, using 4 CPU cores to get 4 MCMC chains 
ret <- ddepn(dataset$datx, phiorig=phit, maxiterations=300, burnin=50,
            plotresults=FALSE, inference="mcmc",
            usebics=FALSE, priortype="laplaceinhib", lambda=0.01, B=B,
            multicores=TRUE, cores=4)
            
## get the signal network containing the significant edges
net <- create_signetwork(ret, alpha=0.05, adj.method="none", plot=FALSE,
 type="wilcox", alternative="one.sided", paired=FALSE, ord=NULL)

## compare inferred net and prior
par(mfrow=c(1,2))
plotdetailed(net, main="inferred", stimuli=ret$samplings[[1]]$stimuli )
plotdetailed(ret$samplings[[1]]$phi.orig, main="prior", 
			stimuli=ret$samplings[[1]]$stimuli )

## use the LOO crossvalidation for finding edges
net2 <- create_signetwork_cv(ret, alpha=0.05, adj.method="none", 
					plot=FALSE, type="wilcox", alternative="one.sided", 
					paired=FALSE, ord=NULL, sel_policy="strict")


## End(Not run)

ddepn documentation built on May 2, 2019, 4:42 p.m.