netga: netga - Perform network search using a genetic algorithm.

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

Description

Uses a genetic algorithm for network search. Combines selection of best models, crossing over of nodes and connected edges between pairs of nodes and mutation of edges to sample the search space.

Usage

netga(dat, stimuli, P=NULL, maxiterations=1000, p=100,
		q=0.3, m=0.8, hmmiterations=30, multicores=FALSE, usebics=FALSE, 
		cores=2, lambda=NULL, B=NULL, Z=NULL, scorefile=NULL, fanin = 4,
		gam=NULL,it=NULL,K=NULL,quantL=.5,quantBIC=.5,priortype="none", 
		plotresults=TRUE, scale_lik=FALSE, allow.stim.off=TRUE,
		debug=0,retobj=NULL, implementation="C")

Arguments

dat	The data matrix. See ddepn
stimuli	The list of stimuli. See ddepn
P	A list containing a population of network models. If NULL, an initial population of networks is sampled including the empty and fully connected network.
maxiterations	Integer, maximal number of generations.
p	The size of the population. Used only if P is not provided.
q	Double \in [0;1]. The rate of selection or crossing over. At most (1-q) \cdot |P| nodes are selected to persist into the next generation, but only if the fitness score is more optimal than the median of scores in the population. If less than (1-q)*|P| fulfill this criterion, the number q \cdot |P| of crossing overs is increased accordingly.
m	Double \in [0;1]. The mutation rate - m \cdot |P| networks out of the Population are mutated after selection and crossing over. Mutations happen for edges and change the type of the edge randomly.
hmmiterations	Maximum number of iterations during the HMM.
multicores	Use multiple processors.
usebics	Use BIC statistic for model selection.
cores	Number of cores to use in case of multicores=TRUE
lambda	The Prior influence strength in the laplace prior.
B	The Prior information matrix. Corresponds to prior edge probabilities in the final network.
Z	The normalisation factor for the prior distribution.
fanin	Integer: maximal indegree for nodes.
gam	Prior influence strength in scalefree prior. Used as exponent in laplaceinhib prior: see prior for details.
it	Number of iterations to generate background distribution in scalefree prior.
K	Proportionality factor in scalefree prior
quantL	Quantile of Population likelihood. Defines the selection threshold in netga. Also used as posterior selection quantile. Note that the Likelihood or Posterior have to be maximised, so all networks with a likelihood/posterior greater than the threshold are selected.
quantBIC	Quantile of Population likelihood. Defines the selection threshold in netga. Note that the BIC is minimised, so all networks with BIC less than the threshold are selected.
scorefile	Name of pdf-file where intermediate traces of the BIC, the differences of BICs in two succeeding generations and a comparison of optimum and average fitness are drawn. If NULL, the plot is drawn on the standard device for plotting.
priortype	Character. One of none, laplaceinhib, laplace or scalefree for use of the respective prior type. Ignored if usebics=TRUE. For netga, usebics=FALSE, priortype="none" means optimising the likelihood directly, for mcmc_ddepn, none is not allowed.
plotresults	Boolean. If TRUE, some statistics are plotted while netga is running.
scale_lik	Boolean. Perform scaling of the likelihood according to how many data points were used to calculate the overall likelihood.
allow.stim.off	Boolean. If TRUE, the stimulus can become passive at some time. This will generate additional reachable system states, in particular all states from the normal state matrix, generated by the propagation, but with the stimulus node set to 0.
debug	Numeric. If 0, a status bar indicates the progress of the algorithm. If 1 or 2, extra information is printed to the console (for debug=2 more information than for debug=1).
retobj	List. The output generated during a GA run (see function ddepn for argument inference="netga"). Passed by function resume_ddepn to resume the inference.
implementation	String. One of "C","R","R_globalest","C_globalest". Different implementations of the HMM in perform.hmmsearch. If "R", the original pure R-implementation is used, if "C", a ported C-implementation is used. If "R_globalest", an experimental version of the parameter estimation is used in the HMM, "C_globalest" is the C-port of this version. See details for a description.

Details

Usually this function is called internally by ddepn.

Value

A list containing the following objects:
A list named P holding the model objects for each candidate network in the population. Each entry in the list contains the following elements:

phi	Adjacency matrix. The inferred network.
L	Likelihood.
aic	Akaikes Information Criterion.
bic	Bayesian Information Criterion.
posterior	Posterior probability. Only present if usebics=FALSE, i.e. when a prior model is used during inference.
dat	The data matrix.
theta	The parameter matrix for the Gaussians.
gamma	The state transition matrix.
gammaposs	The theoretical state transition matrix, as generated by the effect propagation.
tps	A list. Each element is a vector of time points for each experiment in the data matrix.
stimuli	List of stimuli.
reps	Number of replicates for each experiment.
hmmiterations	Maximum number of iterations during an HMM run.

lastmove	String. Type of the last change that was performed.
coords	Position in the network where the last change was performed.
lambda	Laplace prior hyperparameter.
B	The Prior information matrix. See prior for details.
Z	Laplace prior normalisation factor. (Not used at the moment.)
fanin	Maximal indegree for nodes.
gam	Sparsity prior hyperparameter.
it	Sparsity prior iterations.
K	Sparsity prior scaling factor.
priortype	The prior that was used.

A matrix named scorestats holding in each column the traces of the following statistics, each row corresponds to one iteration:

dL_total	Median of total Likelihood change.
dP_total	Median of total Prior change.
dL_crossover	Median of Likelihood change for crossover.
dL_mutation	Median of Likelihood change for mutation.
dP_crossover	Median of Prior change for crossover.
dP_mutation	Median of Prior change for mutation.
dL_total_abs	As above, but absolute values.
dP_total_abs	As above, but absolute values.
dL_crossover_abs	As above, but absolute values.
dL_mutation_abs	As above, but absolute values.
dP_crossover_abs	As above, but absolute values.
dP_mutation_abs	As above, but absolute values.
score	Trace of the total median score. The score depends on the type of inference, if usebics=TRUE it is the BIC score, if usebics=FALSE and priortype="none", it is the Liklihood score, otherwise the posterior score.

Note

TODO

Author(s)

Christian Bender

References

TODO

See Also

ddepn

Examples

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

## sample a network
n <- 6
signet <- signalnetwork(n=n, nstim=2, cstim=0, prop.inh=0.2)
phit <- signet$phi
stimuli <- signet$stimuli

## sample data
dataset <- makedata(phit, stimuli, mu.bg=1200, sd.bg=400, mu.signal.a=2000, sd.signal.a=1000)

## now the GA
ret <- netga(dataset$dat, stimuli, P=NULL, maxiterations=30, p=15,
    q=0.3, m=0.8, hmmiterations=100, multicores=FALSE, usebics=TRUE, cores=1,
    lambda=NULL, B=NULL, scorefile=NULL, priortype="none")
## now the GA with prior
## use original network as prior matrix
## reset all entries for inhibiting edges 
## to -1
B <- phit
B[B==2] <- -1
ret <- netga(dataset$dat, stimuli, P=NULL, maxiterations=30, p=15,
    q=0.3, m=0.8, hmmiterations=100, multicores=FALSE, usebics=FALSE, cores=1,
    lambda=0.01, B=B, scorefile=NULL, priortype="laplaceinhib")



## End(Not run)

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