| Title: | Estimation of genetic and molecular regulatory networks from high-throughput genomics data |
|---|---|
| Description: | Estimate gene and eQTL networks from high-throughput expression and genotyping assays. |
| Authors: | Robert Castelo [aut, cre], Alberto Roverato [aut] |
| Maintainer: | Robert Castelo <[email protected]> |
| License: | GPL (>= 2) |
| Version: | 2.39.0 |
| Built: | 2024-07-19 10:45:43 UTC |
| Source: | https://github.com/bioc/qpgraph |
Estimate gene and eQTL networks from high-throughput expression and genotyping assays.
qpNrr estimates non-rejection rates for every pair
of variables.
qpAvgNrr estimates average non-rejection rates for
every pair of variables.
qpGenNrr estimates generalized average non-rejection rates
for every pair of variables.
qpEdgeNrr estimate the non-rejection rate of one
pair of variables.
qpCItest performs a conditional independence test
between two variables given a conditioning set.
qpHist plots the distribution of non-rejection rates.
qpGraph obtains a qp-graph from a matrix of
non-rejection rates.
qpAnyGraph obtains an undirected graph from a matrix of
pairwise measurements.
qpGraphDensity calculates and plots the graph density
as function of the non-rejection rate.
qpCliqueNumber calculates the size of the largest
maximal clique (the so-called clique number or maximum clique size) in
a given undirected graph.
qpClique calculates and plots the size of the largest
maximal clique (the so-called clique number or maximum clique size)
as function of the non-rejection rate.
qpGetCliques finds the set of (maximal) cliques of
a given undirected graph.
qpRndWishart random generation for the Wishart
distribution.
qpCov calculates the sample covariance matrix, just as
the function cov() but returning a dspMatrix-class
object which efficiently stores such a dense symmetric matrix.
qpG2Sigma builds a random covariance matrix from an
undrected graph. The inverse of the resulting matrix contains zeroes
at the missing edges of the given undirected graph.
qpUnifRndAssociation builds a matrix of uniformly random
association values between -1 and +1 for all pairs of variables that
follow from the number of variables given as input argument.
qpK2ParCor obtains the partial correlation coefficients
from a given concentration matrix.
qpIPF performs maximum likelihood estimation of a
sample covariance matrix given the independence constraints from
an input list of (maximal) cliques.
qpPAC estimates partial correlation coefficients and
corresponding P-values for each edge in a given undirected graph,
from an input data set.
qpPCC estimates pairwise Pearson correlation coefficients
and their corresponding P-values between all pairs of variables from an
input data set.
qpRndGraph builds a random undirected graph with a
bounded maximum connectivity degree on every vertex.
qpPrecisionRecall calculates the precision-recall curve
for a given measure of association between all pairs of variables in a
matrix.
qpPRscoreThreshold calculates the score threshold at a
given precision or recall level from a given precision-recall curve.
qpFunctionalCoherence estimates functional coherence of
a given transcriptional regulatory network using Gene Ontology
annotations.
qpTopPairs reports a top number of pairs of variables
according to either an association measure and/or occurring in a given
reference graph.
qpPlotNetwork plots a network using the Rgraphviz library.
This package provides an implementation of the procedures described in (Castelo
and Roverato, 2006, 2009) and (Tur, Roverato and Castelo, 2014).
An example of its use for reverse-engineering of
transcriptional regulatory networks from microarray data is available in the
vignette qpTxRegNet and, the same directory, contains a pre-print of a book chapter describing the basic functionality of the package which serves the purpose of a basic users's guide. This package is a contribution to the Bioconductor
(Gentleman et al., 2004) and gR (Lauritzen, 2002) projects.
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n. J. Mach. Learn. Res., 7:2621-2650, 2006.
Castelo, R. and Roverato, A. Reverse engineering molecular regulatory networks from microarray data with qp-graphs. J. Comput. Biol. 16(2):213-227, 2009.
Gentleman, R.C., Carey, V.J., Bates, D.M., Bolstad, B., Dettling, M., Dudoit, S., Ellis, B., Gautier, L., Ge, Y., Gentry, J., Hornik, K. Hothorn, T., Huber, W., Iacus, S., Irizarry, R., Leisch, F., Li, C., Maechler, M. Rosinni, A.J., Sawitzki, G., Smith, C., Smyth, G., Tierney, L., Yang, T.Y.H. and Zhang, J. Bioconductor: open software development for computational biology and bioinformatics. Genome Biol., 5:R80, 2004.
Lauritzen, S.L. gRaphical Models in R. R News, 3(2)39, 2002.
Tur, I., Roverato, A. and Castelo, R. Mapping eQTL networks with mixed graphical Markov models. Genetics, 198:1377-1393, 2014.
The data consist of two classes of objects, one containing normalized gene expression microarray data from Escherichia coli (E. coli) and the other containing a subset of filtered RegulonDB transcription regulatory relationships on E. coli.
data(EcoliOxygen)data(EcoliOxygen)
gds680.eset
|
ExpressionSet object containing n=43 experiments
of various mutants under oxygen deprivation (Covert et al., 2004). The
mutants were designed to monitor the response from E. coli during an
oxygen shift in order to target the a priori most relevant part of the
transcriptional network by using six strains with knockouts of five
key transcriptional regulators in the oxygen response (arcA,
appY, fnr, oxyR and soxS). The data was
obtained by downloading the corresponding CEL files from the Gene
Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo) under
accession GDS680 and then normalized using the rma()
function from the affy package. Following the steps described in
(Castelo and Roverato, 2009) probesets were mapped to Entrez Gene
Identifiers and filtered such that the current ExpressionSet
object contains a total of p=4205 genes. The slot featureNames
has already the corresponding Entrez Gene IDs. |
subset.gds680.eset
|
ExpressionSet object corresponding to
a subset of gds680.eset defined by the transcription factor
genes that were knocked-out in the experiments by Covert et al. (2004)
and their putative targets according to the RegulonDB database version
6.1. |
filtered.regulon6.1
|
Data frame object containing a subset of the E. coli transcriptional network from RegulonDB 6.1 (Gama-Castro et al, 2008) obtained through the filtering steps described in (Castelo and Roverato, 2009). In this data frame each row corresponds to a transcriptional regulatory relationship and the first two columns contain Blattner IDs of the transcription factor and target genes, respectively, and the following two correspond to the same genes but specified by Entrez Gene IDs. The fifth column contains the direction of the regulation according to RegulonDB. |
subset.filtered.regulon6.1
|
Subset of filtered.regulon6.1
containing the transcriptional regulatory relationships in RegulonDB
version 6.1 that involve the transcription factor genes which were
knocked-out in the experiments by Covert et al. (2004). |
Covert, M.W., Knight, E.M., Reed, J.L., Herrgard, M.J., and Palsson, B.O. Integrating high-throughput and computational data elucidates bacterial networks. Nature, 429(6987):92-96, 2004.
Gama-Castro, S., Jimenez-Jacinto, V., Peralta-Gil, M., Santos-Zavaleta, A., Penaloza-Spinola, M.I., Contreras-Moreira, B., Segura-Salazar, J., Muniz-Rascado, L., Martinez-Flores, I., Salgado, H., Bonavides-Martinez, C., Abreu-Goodger, C., Rodriguez-Penagos, C., Miranda-Rios, J., Morett, E., Merino, E., Huerta, A.M., Trevino-Quintanilla, L., and Collado-Vides, J. RegulonDB (version 6.0): gene regulation model of Escherichia coli K-12 beyond transcription, active (experimental) annotated promoters and Textpresso navigation. Nucleic Acids Res., 36(Database issue):D120-124, 2008.
Castelo, R. and Roverato, A. Reverse engineering molecular regulatory networks from microarray data with qp-graphs. J. Comp. Biol., 16(2):213-227, 2009.
data(EcoliOxygen) ls()data(EcoliOxygen) ls()
The expression quantitative trait loci (eQTL) experimental cross model class serves the purpose of holding all necessary information to simulate genetical genomics data from an experimental cross.
R. Castelo
The expression quantitative trait loci (eQTL) network class serves the purpose of holding the result of estimating an eQTL network from genetical genomics data.
R. Castelo
eQTLnetworkEstimationParam-class
eQTLcross-class
The methods eQTLnetworkEstimate() perform the estimation of eQTL
networks from genetical genomics data using as input an
eQTLnetworkEstimationParam object and, eventually, a
eQTLnetwork-class object.
R. Castelo
eQTLnetworkEstimationParam
eQTLnetwork-class
The expression quantitative trait loci (eQTL) network parameter class serves
the purpose of holding the parameter information to estimate
an eQTL network from genetical genomics data with the function
eQTLnetworkEstimate.
R. Castelo
eQTLnetworkEstimate
eQTLnetwork-class
Filters out variables or features that lead to collinearities in the input data.
filterCollinearities(X, soft.filter=FALSE, long.dim.are.variables=TRUE)filterCollinearities(X, soft.filter=FALSE, long.dim.are.variables=TRUE)
X |
data set where collinearities are identified. |
soft.filter |
logical; if FALSE (default) then the intput object
|
long.dim.are.variables |
logical; if TRUE (default) it is assumed
that when |
The input object X can be either a matrix object, a
data.frame object or any other class of object that can be
handled by the function qpPCC(), which is internally called,
such as an ExpressionSet object.
The input object X without the variables or features that
lead to collinearities when soft.filter=FALSE, its default
value. Otherwise, when soft.filter=TRUE then a logical mask
is returned.
R. Castelo
## build an undirected GMM model with ## average correlation 0.99 on the present edges set.seed(1234) gmm <- rUGgmm(dRegularGraphParam(), rho=0.99) gmm ## sample n=100 observations from this GMM X <- rmvnorm(100, gmm) dim(X) head(X) ## notice some variables lead to collinearities (r > 0.99) cor(X) ## mask those variables mask <- filterCollinearities(X, long.dim.are.variables=FALSE, soft.filter=TRUE) mask head(X[, !mask])## build an undirected GMM model with ## average correlation 0.99 on the present edges set.seed(1234) gmm <- rUGgmm(dRegularGraphParam(), rho=0.99) gmm ## sample n=100 observations from this GMM X <- rmvnorm(100, gmm) dim(X) head(X) ## notice some variables lead to collinearities (r > 0.99) cor(X) ## mask those variables mask <- filterCollinearities(X, long.dim.are.variables=FALSE, soft.filter=TRUE) mask head(X[, !mask])
Graph parameter classes are defined to ease the simulation of different
types of graphs by using a single interface rgraphBAM().
R. Castelo
The "HMgmm" class is the class of homogeneous mixed graphical
Markov models defined within the qpgraph package to store
simulate and manipulate this type of graphical Markov models (GMMs).
An homogeneous mixed GMM is a family of multivariate conditional Gaussian distributions on mixed discrete and continuous variables sharing a set of conditional independences encoded by means of a marked graph. Further details can be found in the book of Lauritzen (1996).
Objects can be created by calls of the form HMgmm(g, ...) corresponding
to constructor methods or rHMgmm(n, g, ...) corresponding to random
simulation methods.
pI:Object of class "integer" storing the number of
discrete random variables.
pY:Object of class "integer" storing the number of
continuous random variables.
g:Object of class graphBAM-class storing
the associated marked graph.
vtype:Object of class "factor" storing the type (discrete
or continuous) of each random variable.
dLevels:Object of class "integer" storing the number of
levels of each discrete random variable.
a:Object of class "numeric" storing the vector of additive
linear effects on continuous variables connected to discrete ones.
rho:Object of class "numeric" storing the value of the
marginal correlation between two continuous random variables.
sigma:Object of class dspMatrix-class
storing the covariance matrix.
mean:Object of class "numeric" storing the mean vector.
eta2:Object of class "numeric" storing for each continuous
variable connected to a discrete one, the fraction of variance of the
continuous variable explained by the discrete one.
HMgmm(g)Constructor method where g can be either an
adjacency matrix or a graphBAM-class object.
rHMgmm(n, g)Constructor simulation method that allows one to
simulate homogeneous mixed GMMs where n is the number of GMMs to
simulate and g can be either a markedGraphParam object,
an adjacency matrix or a graphBAM-class object.
names(x)Accessor method to obtain the names of the
elements in the object x that can be retrieved with the $
accessor operator.
$Accessor operator to retrieve elements of the object
in an analogous way to a list.
dim(x)Dimension of the homogeneous mixed GMM corresponding to the number of discrete and continuous random variables.
dimnames(x)Names of the discrete and continuous random variables in the homogeneous mixed GMM.
show(object)Method to display some bits of information about
the input homogeneous mixed GMM specified in object.
summary(object)Method to display a sumamry of the main features
of the input homogeneous mixed GMM specified in object.
plot(x, ...)Method to plot the undirected graph associated to the
the input homogeneous mixed GMM specified in x. It uses the plotting
capabilities from the Rgraphviz library to which further arguments
specified in ... are further passed.
R. Castelo
Lauritzen, S.L. Graphical models. Oxford University Press, 1996.
Performs a test of conditional independence for every pair of variables.
## S4 method for signature 'matrix' qpAllCItests(X, I=NULL, Q=NULL, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, return.type=c("p.value", "statn", "all"), verbose=TRUE, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10)## S4 method for signature 'matrix' qpAllCItests(X, I=NULL, Q=NULL, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, return.type=c("p.value", "statn", "all"), verbose=TRUE, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10)
X |
data set from where to estimate the non-rejection rates. It can be an ExpressionSet object, a data frame or a matrix. |
I |
indexes or names of the variables in |
Q |
indexes or names of the variables in |
pairup.i |
subset of vertices to pair up with subset |
pairup.j |
subset of vertices to pair up with subset |
long.dim.are.variables |
logical; if |
exact.test |
logical; if |
use |
a character string defining the way in which calculations are done in the
presence of missing values. It can be either |
tol |
maximum tolerance controlling the convergence of the EM algorithm employed
when the argument |
return.type |
type of value returned by this function. By default |
verbose |
show progress on the calculations. |
R.code.only |
logical; if |
clusterSize |
size of the cluster of processors to employ if we wish to
speed-up the calculations by performing them in parallel. A value of 1
(default) implies a single-processor execution. The use of a cluster of
processors requires having previously loaded the packages |
estimateTime |
logical; if |
nAdj2estimateTime |
number of adjacencies to employ when estimating the
time of calculations ( |
When I is set different to NULL then mixed graphical model theory
is employed and, concretely, it is assumed that the data comes from an homogeneous
conditional Gaussian distribution. By default, with exact.test=TRUE, an
exact test for conditional independence is employed, otherwise an asymptotic one
will be used. Full details on these features can be found in Tur, Roverato and Castelo (2014).
A list with three entries called p.value, statistic and n
corresponding to a dspMatrix-class symmetric matrix of p-values for the null
hypothesis of coindtional independence with the diagonal set to NA values,
an analogous matrix of the statistics of each test and of the sample sizes, respectively.
These returned values, however, depend on the setting of argument return.type which,
by default, enables only returning the matrix of p-values.
If arguments pairup.i and pairup.j are employed, those cells outside
the constrained pairs will get also a NA value.
Note, however, that when estimateTime=TRUE, then instead of the matrix
of estimated non-rejection rates, a vector specifying the estimated number of
days, hours, minutes and seconds for completion of the calculations is returned.
R. Castelo, A. Roverato and I. Tur
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n, J. Mach. Learn. Res., 7:2621-2650, 2006.
Tur, I., Roverato, A. and Castelo, R. Mapping eQTL networks with mixed graphical Markov models. Genetics, 198:1377-1393, 2014.
library(mvtnorm) nVar <- 50 ## number of variables maxCon <- 3 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) alltests <- qpAllCItests(X, verbose=FALSE) ## distribution of p-values for the present edges summary(alltests$p.value[upper.tri(alltests$p.value) & A]) ## distribution of p-values for the missing edges summary(alltests$p.value[upper.tri(alltests$p.value) & !A])library(mvtnorm) nVar <- 50 ## number of variables maxCon <- 3 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) alltests <- qpAllCItests(X, verbose=FALSE) ## distribution of p-values for the present edges summary(alltests$p.value[upper.tri(alltests$p.value) & A]) ## distribution of p-values for the missing edges summary(alltests$p.value[upper.tri(alltests$p.value) & !A])
Obtains an undirected graph from a matrix of pairwise measurements
qpAnyGraph(measurementsMatrix, threshold=NA_real_, remove=c("below", "above"), topPairs=NA_integer_, decreasing=TRUE, pairup.i=NULL, pairup.j=NULL)qpAnyGraph(measurementsMatrix, threshold=NA_real_, remove=c("below", "above"), topPairs=NA_integer_, decreasing=TRUE, pairup.i=NULL, pairup.j=NULL)
measurementsMatrix |
matrix of pairwise measurements. |
threshold |
threshold on the measurements below or above which pairs of variables are assumed to be disconnected in the resulting graph. |
remove |
direction of the removal with the threshold. It should be
either |
topPairs |
number of edges from the top of the ranking, defined by the
pairwise measurements in |
decreasing |
logical, only applies when topPairs is set; if |
pairup.i |
subset of vertices to pair up with subset |
pairup.j |
subset of vertices to pair up with subset |
This is a general purpose function for thresholding a matrix of pairwise measurements to select pairs of variables corresponding to selected edges in an undirected graph.
The resulting undirected graph as a graphBAM object.
Note that when some gold-standard graph is available for comparison,
a value for the parameter threshold can be found by calculating a
precision-recall curve with qpPrecisionRecall with respect to this
gold-standard, and then using qpPRscoreThreshold. Parameters
threshold and topPairs are mutually exclusive, that is, when
we specify with topPairs=n that we want a graph with n edges
then threshold cannot be used.
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n, J. Mach. Learn. Res., 7:2621-2650, 2006.
qpNrr
qpAvgNrr
qpEdgeNrr
qpGraph
qpGraphDensity
qpClique
qpPrecisionRecall
qpPRscoreThreshold
require(mvtnorm) require(graph) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## estimate Pearson correlations pcc.estimates <- qpPCC(X) ## the higher the threshold g <- qpAnyGraph(abs(pcc.estimates$R), threshold=0.9, remove="below") ## the sparser the qp-graph numEdges(g) / choose(numNodes(g), 2) ## the lower the threshold g <- qpAnyGraph(abs(pcc.estimates$R), threshold=0.5, remove="below") # the denser the graph numEdges(g) / choose(numNodes(g), 2)require(mvtnorm) require(graph) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## estimate Pearson correlations pcc.estimates <- qpPCC(X) ## the higher the threshold g <- qpAnyGraph(abs(pcc.estimates$R), threshold=0.9, remove="below") ## the sparser the qp-graph numEdges(g) / choose(numNodes(g), 2) ## the lower the threshold g <- qpAnyGraph(abs(pcc.estimates$R), threshold=0.5, remove="below") # the denser the graph numEdges(g) / choose(numNodes(g), 2)
Estimates average non-rejection rates for every pair of variables.
## S4 method for signature 'ExpressionSet' qpAvgNrr(X, qOrders=4, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, type=c("arith.mean"), verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'data.frame' qpAvgNrr(X, qOrders=4, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, type=c("arith.mean"), verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'matrix' qpAvgNrr(X, qOrders=4, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, type=c("arith.mean"), verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10)## S4 method for signature 'ExpressionSet' qpAvgNrr(X, qOrders=4, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, type=c("arith.mean"), verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'data.frame' qpAvgNrr(X, qOrders=4, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, type=c("arith.mean"), verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'matrix' qpAvgNrr(X, qOrders=4, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, type=c("arith.mean"), verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10)
X |
data set from where to estimate the average non-rejection rates. It can be an ExpressionSet object, a data frame or a matrix. |
qOrders |
either a number of partial-correlation orders or a vector of vector of particular orders to be employed in the calculation. |
I |
indexes or names of the variables in |
restrict.Q |
indexes or names of the variables in |
fix.Q |
indexes or names of the variables in |
nTests |
number of tests to perform for each pair for variables. |
alpha |
significance level of each test. |
pairup.i |
subset of vertices to pair up with subset |
pairup.j |
subset of vertices to pair up with subset |
long.dim.are.variables |
logical; if |
type |
type of average. By now only the arithmetic mean is available. |
verbose |
show progress on the calculations. |
identicalQs |
use identical conditioning subsets for every pair of vertices
(default), otherwise sample a new collection of |
exact.test |
logical; if |
use |
a character string defining the way in which calculations are done in the
presence of missing values. It can be either |
tol |
maximum tolerance controlling the convergence of the EM algorithm employed
when the argument |
R.code.only |
logical; if |
clusterSize |
size of the cluster of processors to employ if we wish to
speed-up the calculations by performing them in parallel. A value of 1
(default) implies a single-processor execution. The use of a cluster of
processors requires having previously loaded the packages |
estimateTime |
logical; if |
nAdj2estimateTime |
number of adjacencies to employ when estimating the
time of calculations ( |
Note that when specifying a vector of particular orders q, these values
should be in the range 1 to min(p, n-3), where p is the number of
variables and n the number of observations. The computational cost
increases linearly within each q value and quadratically in p.
When setting identicalQs to FALSE the computational cost may
increase between 2 times and one order of magnitude (depending on p and
q) while asymptotically the estimation of the non-rejection rate
converges to the same value.
When I is set different to NULL then mixed graphical model theory
is employed and, concretely, it is assumed that the data comes from an homogeneous
conditional Gaussian distribution. In this setting further restrictions to the
maximum value of q apply, concretely, it cannot be smaller than
p plus the number of levels of the discrete variables involved in the
marginal distributions employed by the algorithm. By default, with
exact.test=TRUE, an exact test for conditional independence is employed,
otherwise an asymptotic one will be used. Full details on these features can
be found in Tur, Roverato and Castelo (2014).
A dspMatrix-class symmetric matrix of estimated average
non-rejection rates with the diagonal set to NA values. When using the
arguments pairup.i and pairup.j, those cells outside the
constraint pairs will get also a NA value.
Note, however, that when estimateTime=TRUE, then instead of the matrix
of estimated average non-rejection rates, a vector specifying the estimated
number of days, hours, minutes and seconds for completion of the calculations
is returned.
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. Reverse engineering molecular regulatory networks from microarray data with qp-graphs. J. Comp. Biol., 16(2):213-227, 2009.
Tur, I., Roverato, A. and Castelo, R. Mapping eQTL networks with mixed graphical Markov models. Genetics, 198:1377-1393, 2014.
qpNrr
qpEdgeNrr
qpHist
qpGraphDensity
qpClique
require(mvtnorm) nVar <- 50 ## number of variables maxCon <- 3 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) avgnrr.estimates <- qpAvgNrr(X, verbose=FALSE) ## distribution of average non-rejection rates for the present edges summary(avgnrr.estimates[upper.tri(avgnrr.estimates) & A]) ## distribution of average non-rejection rates for the missing edges summary(avgnrr.estimates[upper.tri(avgnrr.estimates) & !A]) ## Not run: library(snow) library(rlecuyer) ## only for moderate and large numbers of variables the ## use of a cluster of processors speeds up the calculations nVar <- 500 maxCon <- 3 A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) system.time(avgnrr.estimates <- qpAvgNrr(X, q=10, verbose=TRUE)) system.time(avgnrr.estimates <- qpAvgNrr(X, q=10, verbose=TRUE, clusterSize=4)) ## End(Not run)require(mvtnorm) nVar <- 50 ## number of variables maxCon <- 3 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) avgnrr.estimates <- qpAvgNrr(X, verbose=FALSE) ## distribution of average non-rejection rates for the present edges summary(avgnrr.estimates[upper.tri(avgnrr.estimates) & A]) ## distribution of average non-rejection rates for the missing edges summary(avgnrr.estimates[upper.tri(avgnrr.estimates) & !A]) ## Not run: library(snow) library(rlecuyer) ## only for moderate and large numbers of variables the ## use of a cluster of processors speeds up the calculations nVar <- 500 maxCon <- 3 A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) system.time(avgnrr.estimates <- qpAvgNrr(X, q=10, verbose=TRUE)) system.time(avgnrr.estimates <- qpAvgNrr(X, q=10, verbose=TRUE, clusterSize=4)) ## End(Not run)
Calculates and plots the size of the largest vertex boundary as function of the non-rejection rate.
qpBoundary(nrrMatrix, n=NA, threshold.lim=c(0,1), breaks=5, vertexSubset=NULL, plot=TRUE, qpBoundaryOutput=NULL, density.digits=0, logscale.bdsize=FALSE, titlebd="Maximum boundary size as function of threshold", verbose=FALSE)qpBoundary(nrrMatrix, n=NA, threshold.lim=c(0,1), breaks=5, vertexSubset=NULL, plot=TRUE, qpBoundaryOutput=NULL, density.digits=0, logscale.bdsize=FALSE, titlebd="Maximum boundary size as function of threshold", verbose=FALSE)
nrrMatrix |
matrix of non-rejection rates. |
n |
number of observations from where the non-rejection rates were estimated. |
threshold.lim |
range of threshold values on the non-rejection rate. |
breaks |
either a number of threshold bins or a vector of threshold breakpoints. |
vertexSubset |
subset of vertices for which their maximum boundary size is calculated with respect to all other vertices. |
plot |
logical; if TRUE makes a plot of the result; if FALSE it does not. |
qpBoundaryOutput |
output from a previous call to |
density.digits |
number of digits in the reported graph densities. |
logscale.bdsize |
logical; if TRUE then the scale for the maximum boundary size is logarithmic which is useful when working with more than 1000 variables; FALSE otherwise (default). |
titlebd |
main title to be shown in the plot. |
verbose |
show progress on calculations. |
The maximum boundary is calculated as the largest degree among all vertices of a given qp-graph.
A list with the maximum boundary size and graph density as function of threshold, the threshold on the non-rejection rate that provides a maximum boundary size strictly smaller than the sample size n and the resulting maximum boundary size.
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n. J. Mach. Learn. Res., 7:2621-2650, 2006.
require(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## the higher the q the less complex the qp-graph nrr.estimates <- qpNrr(X, q=1, verbose=FALSE) qpBoundary(nrr.estimates, plot=FALSE) nrr.estimates <- qpNrr(X, q=5, verbose=FALSE) qpBoundary(nrr.estimates, plot=FALSE)require(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## the higher the q the less complex the qp-graph nrr.estimates <- qpNrr(X, q=1, verbose=FALSE) qpBoundary(nrr.estimates, plot=FALSE) nrr.estimates <- qpNrr(X, q=5, verbose=FALSE) qpBoundary(nrr.estimates, plot=FALSE)
Performs a conditional independence test between two variables given a conditioning set.
## S4 method for signature 'ExpressionSet' qpCItest(X, i=1, j=2, Q=c(), exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'cross' qpCItest(X, i=1, j=2, Q=c(), exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'data.frame' qpCItest(X, i=1, j=2, Q=c(), I=NULL, long.dim.are.variables=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'matrix' qpCItest(X, i=1, j=2, Q=c(), I=NULL, long.dim.are.variables=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'SsdMatrix' qpCItest(X, i=1, j=2, Q=c(), R.code.only=FALSE)## S4 method for signature 'ExpressionSet' qpCItest(X, i=1, j=2, Q=c(), exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'cross' qpCItest(X, i=1, j=2, Q=c(), exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'data.frame' qpCItest(X, i=1, j=2, Q=c(), I=NULL, long.dim.are.variables=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'matrix' qpCItest(X, i=1, j=2, Q=c(), I=NULL, long.dim.are.variables=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'SsdMatrix' qpCItest(X, i=1, j=2, Q=c(), R.code.only=FALSE)
X |
data set where the test should be performed. It can be either
an |
i |
index or name of one of the two variables in |
j |
index or name of the other variable in |
Q |
indexes or names of the variables in |
I |
indexes or names of the variables in |
long.dim.are.variables |
logical; if TRUE it is assumed that when data are in a data frame or in a matrix, the longer dimension is the one defining the random variables (default); if FALSE, then random variables are assumed to be at the columns of the data frame or matrix. |
exact.test |
logical; if |
use |
a character string defining the way in which calculations are done in the
presence of missing values. It can be either |
tol |
maximum tolerance controlling the convergence of the EM algorithm employed
when the argument |
R.code.only |
logical; if FALSE then the faster C implementation is used (default); if TRUE then only R code is executed. |
When variables in i, j and Q are continuous and I=NULL, this function
performs a conditional independence test using a t-test for zero partial regression coefficient
(Lauritzen, 1996, pg. 150). Note that the size of possible Q sets should be in
the range 1 to min(p,n-3), where p is the number of variables and n
the number of observations. The computational cost increases linearly with
the number of variables in Q.
When variables in i, j and Q are continuous and discrete (mixed data),
indicated with the I argument when X is a matrix, then mixed graphical
model theory (Lauritzen and Wermuth, 1989) is employed and, concretely, it is assumed
that data come from an homogeneous conditional Gaussian distribution. By default, with
exact.test=TRUE, an exact likelihood ratio test for conditional independence is
performed (Lauritzen, 1996, pg. 192-194; Tur, Roverato and Castelo, 2014), otherwise an
asymptotic one is used.
In this setting further restrictions to the maximum value of q apply, concretely,
it cannot be smaller than p plus the number of levels of the discrete variables
involved in the marginal distributions employed by the algorithm.
A list with class "htest" containing the following components:
statistic |
in case of pure continuous data and |
parameter |
in case of pure continuous data and |
p.value |
the p-value for the test. |
estimate |
in case of pure continuous data ( |
alternative |
a character string describing the alternative hypothesis. |
method |
a character string indicating what type of conditional independence test was performed. |
data.name |
a character string giving the name(s) of the random variables involved in the conditional independence test. |
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n, J. Mach. Learn. Res., 7:2621-2650, 2006.
Lauritzen, S.L. Graphical models. Oxford University Press, 1996.
Lauritzen, S.L and Wermuth, N. Graphical Models for associations between variables, some of which are qualitative and some quantitative. Ann. Stat., 17(1):31-57, 1989.
Tur, I., Roverato, A. and Castelo, R. Mapping eQTL networks with mixed graphical Markov models. Genetics, 198:1377-1393, 2014.
require(mvtnorm) nObs <- 100 ## number of observations to simulate ## the following adjacency matrix describes an undirected graph ## where vertex 3 is conditionally independent of 4 given 1 AND 2 A <- matrix(c(FALSE, TRUE, TRUE, TRUE, TRUE, FALSE, TRUE, TRUE, TRUE, TRUE, FALSE, FALSE, TRUE, TRUE, FALSE, FALSE), nrow=4, ncol=4, byrow=TRUE) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) qpCItest(X, i=3, j=4, Q=1, long.dim.are.variables=FALSE) qpCItest(X, i=3, j=4, Q=c(1,2), long.dim.are.variables=FALSE)require(mvtnorm) nObs <- 100 ## number of observations to simulate ## the following adjacency matrix describes an undirected graph ## where vertex 3 is conditionally independent of 4 given 1 AND 2 A <- matrix(c(FALSE, TRUE, TRUE, TRUE, TRUE, FALSE, TRUE, TRUE, TRUE, TRUE, FALSE, FALSE, TRUE, TRUE, FALSE, FALSE), nrow=4, ncol=4, byrow=TRUE) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) qpCItest(X, i=3, j=4, Q=1, long.dim.are.variables=FALSE) qpCItest(X, i=3, j=4, Q=c(1,2), long.dim.are.variables=FALSE)
Calculates and plots the size of the largest maximal clique (the so-called clique number or maximum clique size) as function of the non-rejection rate.
qpClique(nrrMatrix, n=NA, threshold.lim=c(0,1), breaks=5, plot=TRUE, exact.calculation=TRUE, approx.iter=100, qpCliqueOutput=NULL, density.digits=0, logscale.clqsize=FALSE, titleclq="maximum clique size as function of threshold", verbose=FALSE)qpClique(nrrMatrix, n=NA, threshold.lim=c(0,1), breaks=5, plot=TRUE, exact.calculation=TRUE, approx.iter=100, qpCliqueOutput=NULL, density.digits=0, logscale.clqsize=FALSE, titleclq="maximum clique size as function of threshold", verbose=FALSE)
nrrMatrix |
matrix of non-rejection rates. |
n |
number of observations from where the non-rejection rates were estimated. |
threshold.lim |
range of threshold values on the non-rejection rate. |
breaks |
either a number of threshold bins or a vector of threshold breakpoints. |
plot |
logical; if TRUE makes a plot of the result; if FALSE it does not. |
exact.calculation |
logical; if TRUE then the exact clique number is calculated; if FALSE then a lower bound is given instead. |
approx.iter |
number of iterations to be employed in the calculation of
the lower bound (i.e., only applies when |
qpCliqueOutput |
output from a previous call to |
density.digits |
number of digits in the reported graph densities. |
logscale.clqsize |
logical; if TRUE then the scale for the maximum clique size is logarithmic which is useful when working with more than 1000 variables; FALSE otherwise (default). |
titleclq |
main title to be shown in the plot. |
verbose |
show progress on calculations. |
The estimate of the complexity of the resulting qp-graphs is calculated as the area enclosed under the curve of maximum clique sizes.
The maximum clique size, or clique number, is obtained by calling the function
qpCliqueNumber
The calculation of the clique number of an undirected graph is an NP-complete
problem which means that its computational cost is bounded by an exponential
running time (Pardalos and Xue, 1994). Therefore, giving breakpoints between
0.95 and 1.0 may result into very dense graphs which can lead to extremely
long execution times. If it is necessary to look at that range of breakpoints
it is recommended either to use the lower bound on the clique number
(exact.calculation=FALSE) or to look at qpGraphDensity.
A list with the maximum clique size and graph density as function of threshold, an estimate of the complexity of the resulting qp-graphs across the thresholds, the threshold on the non-rejection rate that provides a maximum clique size strictly smaller than the sample size n and the resulting maximum clique size.
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n. J. Mach. Learn. Res., 7:2621-2650, 2006.
Pardalos, P.M. and Xue, J. The maximum clique problem. J. Global Optim., 4:301-328, 1994.
require(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## the higher the q the less complex the qp-graph nrr.estimates <- qpNrr(X, q=1, verbose=FALSE) qpClique(nrr.estimates, plot=FALSE)$complexity nrr.estimates <- qpNrr(X, q=5, verbose=FALSE) qpClique(nrr.estimates, plot=FALSE)$complexityrequire(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## the higher the q the less complex the qp-graph nrr.estimates <- qpNrr(X, q=1, verbose=FALSE) qpClique(nrr.estimates, plot=FALSE)$complexity nrr.estimates <- qpNrr(X, q=5, verbose=FALSE) qpClique(nrr.estimates, plot=FALSE)$complexity
Calculates the size of the largest maximal clique (the so-called clique number or maximum clique size) in a given undirected graph.
qpCliqueNumber(g, exact.calculation=TRUE, return.vertices=FALSE, approx.iter=100, verbose=TRUE, R.code.only)qpCliqueNumber(g, exact.calculation=TRUE, return.vertices=FALSE, approx.iter=100, verbose=TRUE, R.code.only)
g |
either a |
exact.calculation |
logical; if TRUE then the exact clique number is calculated; if FALSE then a lower bound is given instead. |
return.vertices |
logical; if TRUE a set of vertices forming a maximal clique of maximum size is returned; if FALSE only the maximum clique size is returned. |
approx.iter |
number of iterations to be employed in the calculation of
the lower bound (i.e., only applies when |
verbose |
show progress on calculations. |
R.code.only |
logical; if FALSE then the faster C implementation is used (default); if TRUE then only R code is executed. |
The calculation of the clique number of an undirected graph is one of the basic NP-complete problems (Karp, 1972) which means that its computational cost is bounded by an exponential running time (Pardalos and Xue, 1994). The current implementation uses C code from the GNU GPL Cliquer library by Niskanen and Ostergard (2003) based on the, probably the fastest to date, algorithm by Ostergard (2002).
The lower bound on the maximum clique size is calculated by ranking the
vertices by their connectivity degree, put the first vertex in a set and
go through the rest of the ranking adding those vertices to the set that
form a clique with the vertices currently within the set. Once the entire
ranking has been examined a large clique should have been built and eventually
one of the largests ones. This process is repeated a number of times
(approx.iter) each of which the ranking is altered with increasing
levels of randomness acyclically (altering 1 to $p$ vertices and again). Larger
values of approx.iter should provide tighter lower bounds although it has
been proven that no polynomial time algorithm can approximate the maximum
clique size within a factor of (), unless P=NP
(Feige et al, 1991; Pardalos and Xue, 1994).
a lower bound of the size of the largest maximal clique in the given graph, also known as its clique number.
R. Castelo
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n. J. Mach. Learn. Res., 7:2621-2650, 2006.
Feige, U., Goldwasser, S., Lov\'asz, L., Safra, S. and Szegedy, M. Approximating the maximum clique is almost NP-Complete. Proc. 32nd IEEE Symp. on Foundations of Computer Science, 2-12, 1991.
Karp, R.M. Reducibility among combinatorial problems. Complexity of computer computations, 43:85-103, 1972.
Niskanen, S. Ostergard, P. Cliquer User's Guide, Version 1.0. Communications Laboratory, Helsinki University of Technology, Espoo, Finland, Tech. Rep. T48, 2003. (http://users.tkk.fi/~pat/cliquer.html)
Ostergard, P. A fast algorithm for the maximum clique problem. Discrete Appl. Math. 120:197-207, 2002.
Pardalos, P.M. and Xue, J. The maximum clique problem. J. Global Optim., 4:301-328, 1994.
require(graph) nVar <- 50 set.seed(123) g1 <- randomEGraph(V=as.character(1:nVar), p=0.3) qpCliqueNumber(g1, verbose=FALSE) g2 <- randomEGraph(V=as.character(1:nVar), p=0.7) qpCliqueNumber(g2, verbose=FALSE)require(graph) nVar <- 50 set.seed(123) g1 <- randomEGraph(V=as.character(1:nVar), p=0.3) qpCliqueNumber(g1, verbose=FALSE) g2 <- randomEGraph(V=as.character(1:nVar), p=0.7) qpCliqueNumber(g2, verbose=FALSE)
Calculates the sample covariance matrix, just as the function cov()
but returning a dspMatrix-class object which efficiently
stores such a dense symmetric matrix.
qpCov(X, corrected=TRUE)qpCov(X, corrected=TRUE)
X |
data set from where to calculate the sample covariance matrix.
As the |
corrected |
flag set to |
This function makes the same calculation as the cov function
but returns a sample covariance matrix stored in the space-efficient class
dspMatrix-class and, moreover, allows one for calculating
the uncorrected sum of squares and deviations which equals
(n-1) * cov().
A sample covariance matrix stored as a dspMatrix-class object.
See the Matrix package for full details on this object class.
R. Castelo
require(graph) require(mvtnorm) nVar <- 50 ## number of variables nObs <- 10 ## number of observations to simulate set.seed(123) g <- randomEGraph(as.character(1:nVar), p=0.15) Sigma <- qpG2Sigma(g, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) S <- qpCov(X) ## estimate Pearson correlation coefficients by scaling the sample covariance matrix R <- cov2cor(as(S, "matrix")) ## get the corresponding boolean adjacency matrix A <- as(g, "matrix") == 1 ## Pearson correlation coefficients of the present edges summary(abs(R[upper.tri(R) & A])) ## Pearson correlation coefficients of the missing edges summary(abs(R[upper.tri(R) & !A]))require(graph) require(mvtnorm) nVar <- 50 ## number of variables nObs <- 10 ## number of observations to simulate set.seed(123) g <- randomEGraph(as.character(1:nVar), p=0.15) Sigma <- qpG2Sigma(g, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) S <- qpCov(X) ## estimate Pearson correlation coefficients by scaling the sample covariance matrix R <- cov2cor(as(S, "matrix")) ## get the corresponding boolean adjacency matrix A <- as(g, "matrix") == 1 ## Pearson correlation coefficients of the present edges summary(abs(R[upper.tri(R) & A])) ## Pearson correlation coefficients of the missing edges summary(abs(R[upper.tri(R) & !A]))
Estimates the non-rejection rate for one pair of variables.
## S4 method for signature 'ExpressionSet' qpEdgeNrr(X, i=1, j=2, q=1, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'data.frame' qpEdgeNrr(X, i=1, j=2, q=1, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, long.dim.are.variables=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'matrix' qpEdgeNrr(X, i=1, j=2, q=1, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, long.dim.are.variables=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'SsdMatrix' qpEdgeNrr(X, i=1, j=2, q=1, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, R.code.only=FALSE)## S4 method for signature 'ExpressionSet' qpEdgeNrr(X, i=1, j=2, q=1, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'data.frame' qpEdgeNrr(X, i=1, j=2, q=1, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, long.dim.are.variables=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'matrix' qpEdgeNrr(X, i=1, j=2, q=1, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, long.dim.are.variables=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE) ## S4 method for signature 'SsdMatrix' qpEdgeNrr(X, i=1, j=2, q=1, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, R.code.only=FALSE)
X |
data set from where the non-rejection rate should be estimated. It
can be either an |
i |
index or name of one of the two variables in |
j |
index or name of the other variable in |
q |
order of the conditioning subsets employed in the calculation. |
I |
indexes or names of the variables in |
restrict.Q |
indexes or names of the variables in |
fix.Q |
indexes or names of the variables in |
nTests |
number of tests to perform for each pair for variables. |
alpha |
significance level of each test. |
long.dim.are.variables |
logical; if TRUE it is assumed that when data are in a data frame or in a matrix, the longer dimension is the one defining the random variables (default); if FALSE, then random variables are assumed to be at the columns of the data frame or matrix. |
exact.test |
logical; if |
use |
a character string defining the way in which calculations are done in the
presence of missing values. It can be either |
tol |
maximum tolerance controlling the convergence of the EM algorithm employed
when the argument |
R.code.only |
logical; if FALSE then the faster C implementation is used (default); if TRUE then only R code is executed. |
The estimation of the non-rejection rate for a pair of variables is calculated as the fraction of tests that accept the null hypothesis of conditional independence given a set of randomly sampled q-order conditionals.
Note that the possible values of q should be in the range 1 to
min(p,n-3), where p is the number of variables and n
the number of observations. The computational cost increases linearly with
q.
When I is set different to NULL then mixed graphical model theory
is employed and, concretely, it is assumed that the data comes from an homogeneous
conditional Gaussian distribution. In this setting further restrictions to the
maximum value of q apply, concretely, it cannot be smaller than
p plus the number of levels of the discrete variables involved in the
marginal distributions employed by the algorithm. By default, with
exact.test=TRUE, an exact test for conditional independence is employed,
otherwise an asymptotic one will be used. Full details on these features can
be found in Tur, Roverato and Castelo (2014).
The argument I specifying what variables are discrete actually applies only
when X is a matrix object since in the other cases data types are specified
for each data columns or slot.
An estimate of the non-rejection rate for the particular given pair of variables.
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n, J. Mach. Learn. Res., 7:2621-2650, 2006.
Tur, I., Roverato, A. and Castelo, R. Mapping eQTL networks with mixed graphical Markov models. Genetics, 198:1377-1393, 2014.
qpNrr
qpAvgNrr
qpHist
qpGraphDensity
qpClique
qpCov
require(mvtnorm) nObs <- 100 ## number of observations to simulate ## the following adjacency matrix describes an undirected graph ## where vertex 3 is conditionally independent of 4 given 1 AND 2 A <- matrix(c(FALSE, TRUE, TRUE, TRUE, TRUE, FALSE, TRUE, TRUE, TRUE, TRUE, FALSE, FALSE, TRUE, TRUE, FALSE, FALSE), nrow=4, ncol=4, byrow=TRUE) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) qpEdgeNrr(X, i=3, j=4, q=1, long.dim.are.variables=FALSE) qpEdgeNrr(X, i=3, j=4, q=2, long.dim.are.variables=FALSE)require(mvtnorm) nObs <- 100 ## number of observations to simulate ## the following adjacency matrix describes an undirected graph ## where vertex 3 is conditionally independent of 4 given 1 AND 2 A <- matrix(c(FALSE, TRUE, TRUE, TRUE, TRUE, FALSE, TRUE, TRUE, TRUE, TRUE, FALSE, FALSE, TRUE, TRUE, FALSE, FALSE), nrow=4, ncol=4, byrow=TRUE) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) qpEdgeNrr(X, i=3, j=4, q=1, long.dim.are.variables=FALSE) qpEdgeNrr(X, i=3, j=4, q=2, long.dim.are.variables=FALSE)
Estimates functional coherence for a given transcriptional regulatory network specified either as an adjacency matrix with a list of transcription factor gene identifiers or as a list of transcriptional regulatory modules, whose element names determine which genes encode for transcription factor proteins.
## S4 method for signature 'lsCMatrix' qpFunctionalCoherence(object, TFgenes, geneUniverse=rownames(object), chip, minRMsize=5, removeGOterm="transcription", verbose=FALSE, clusterSize=1) ## S4 method for signature 'lspMatrix' qpFunctionalCoherence(object, TFgenes, geneUniverse=rownames(object), chip, minRMsize=5, removeGOterm="transcription", verbose=FALSE, clusterSize=1) ## S4 method for signature 'lsyMatrix' qpFunctionalCoherence(object, TFgenes, geneUniverse=rownames(object), chip, minRMsize=5, removeGOterm="transcription", verbose=FALSE, clusterSize=1) ## S4 method for signature 'matrix' qpFunctionalCoherence(object, TFgenes, geneUniverse=rownames(object), chip, minRMsize=5, removeGOterm="transcription", verbose=FALSE, clusterSize=1) ## S4 method for signature 'list' qpFunctionalCoherence(object, geneUniverse=unique(c(names(object), unlist(object, use.names=FALSE))), chip, minRMsize=5, removeGOterm="transcription", verbose=FALSE, clusterSize=1)## S4 method for signature 'lsCMatrix' qpFunctionalCoherence(object, TFgenes, geneUniverse=rownames(object), chip, minRMsize=5, removeGOterm="transcription", verbose=FALSE, clusterSize=1) ## S4 method for signature 'lspMatrix' qpFunctionalCoherence(object, TFgenes, geneUniverse=rownames(object), chip, minRMsize=5, removeGOterm="transcription", verbose=FALSE, clusterSize=1) ## S4 method for signature 'lsyMatrix' qpFunctionalCoherence(object, TFgenes, geneUniverse=rownames(object), chip, minRMsize=5, removeGOterm="transcription", verbose=FALSE, clusterSize=1) ## S4 method for signature 'matrix' qpFunctionalCoherence(object, TFgenes, geneUniverse=rownames(object), chip, minRMsize=5, removeGOterm="transcription", verbose=FALSE, clusterSize=1) ## S4 method for signature 'list' qpFunctionalCoherence(object, geneUniverse=unique(c(names(object), unlist(object, use.names=FALSE))), chip, minRMsize=5, removeGOterm="transcription", verbose=FALSE, clusterSize=1)
object |
object containing the transcriptional regulatory modules for which we want to estimate their functional coherence. It can be an adjacency matrix of the undirected graph representing the transcriptional regulatory network or a list of gene target sets where the name of the entry should be the transcription factor gene identifier. |
TFgenes |
when the input object is a matrix, it is required to provide a vector of transcription factor gene identifiers (which should match somewhere in the row and column names of the matrix. |
geneUniverse |
vector of all genes considered in the analysis. By default
it equals the rows and column names of |
chip |
name of the |
minRMsize |
minimum size of the target gene set in each regulatory module where functional enrichment will be calculated and thus where functional coherence will be estimated. |
removeGOterm |
word, or regular pattern, matching GO terms that should be excluded in the transcription factor gene GO annotations, and in the target gene if the regulatory module has only one gene, prior to the calculation of functional coherence. |
verbose |
logical; if TRUE the function will show progress on the calculations; if FALSE the function will remain quiet (default). |
clusterSize |
size of the cluster of processors to employ if we wish to
speed-up the calculations by performing them in parallel. A value of 1
(default) implies a single-processor execution. The use of a cluster of
processors requires having previously loaded the packages |
This function estimates the functional coherence of a transcriptional regulatory
network represented by means of an undirected graph encoded by either an adjacency matrix
and a vector of transcription factor genes, or a list of regulatory modules each of
them defined by a transcription factor gene and its targets. The functional coherence of a
transcriptional regulatory network is calculated as specified by Castelo and
Roverato (2009) and corresponds to the distribution of individual functional
coherence values of every of the regulatory modules of the network each of them
defined as a transcription factor and its set of putatively regulated target
genes. In the calculation of the functional coherence value of a regulatory
module, Gene Ontology (GO) annotations are employed through the given annotation
.db package and the conditional hyper-geometric test implemented in the
GOstats package from Bioconductor.
When a regulatory module has only one target gene, then no functional enrichment is calculated and, instead, the GO trees, grown from the GO annotations of the transcription factor gene and its target, are directly compared.
A list with the following elements: the transcriptional regulatory network as a list of regulatory modules and their targets; the previous list of regulatory modules but excluding those with no enriched GO BP terms. When the regulatory module has only one target, then instead the GO BP annotations of the target gene are included; a vector of functional coherence values.
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. Reverse engineering molecular regulatory networks from microarray data with qp-graphs. J. Comp. Biol., 16(2):213-227, 2009.
## example below takes about minute and a half to execute and for ## that reason it is not executed by default ## Not run: library(GOstats) library(org.EcK12.eg.db) ## load RegulonDB data from this package data(EcoliOxygen) ## pick two TFs from the RegulonDB data in this package TFgenes <- c("mhpR", "iscR") ## get their Entrez Gene Identifiers TFgenesEgIDs <- unlist(mget(TFgenes, AnnotationDbi::revmap(org.EcK12.egSYMBOL))) ## get all genes involved in their regulatory modules from ## the RegulonDB data in this package mt <- match(filtered.regulon6.1[,"EgID_TF"], TFgenesEgIDs) allGenes <- as.character(unique(as.vector( as.matrix(filtered.regulon6.1[!is.na(mt), c("EgID_TF","EgID_TG")])))) mtTF <- match(filtered.regulon6.1[,"EgID_TF"],allGenes) mtTG <- match(filtered.regulon6.1[,"EgID_TG"],allGenes) ## select the corresponding subset of the RegulonDB data in this package subset.filtered.regulon6.1 <- filtered.regulon6.1[!is.na(mtTF) & !is.na(mtTG),] TFi <- match(subset.filtered.regulon6.1[,"EgID_TF"], allGenes) TGi <- match(subset.filtered.regulon6.1[,"EgID_TG"], allGenes) subset.filtered.regulon6.1 <- cbind(subset.filtered.regulon6.1, idx_TF=TFi, idx_TG=TGi) ## build an adjacency matrix representing the transcriptional regulatory ## relationships from these regulatory modules p <- length(allGenes) adjacencyMatrix <- matrix(FALSE, nrow=p, ncol=p) rownames(adjacencyMatrix) <- colnames(adjacencyMatrix) <- allGenes idxTFTG <- as.matrix(subset.filtered.regulon6.1[,c("idx_TF","idx_TG")]) adjacencyMatrix[idxTFTG] <- adjacencyMatrix[cbind(idxTFTG[,2],idxTFTG[,1])] <- TRUE ## calculate functional coherence on these regulatory modules fc <- qpFunctionalCoherence(adjacencyMatrix, TFgenes=TFgenesEgIDs, chip="org.EcK12.eg.db") print(sprintf("the %s module has a FC value of %.2f", mget(names(fc$functionalCoherenceValues),org.EcK12.egSYMBOL), fc$functionalCoherenceValues)) ## End(Not run)## example below takes about minute and a half to execute and for ## that reason it is not executed by default ## Not run: library(GOstats) library(org.EcK12.eg.db) ## load RegulonDB data from this package data(EcoliOxygen) ## pick two TFs from the RegulonDB data in this package TFgenes <- c("mhpR", "iscR") ## get their Entrez Gene Identifiers TFgenesEgIDs <- unlist(mget(TFgenes, AnnotationDbi::revmap(org.EcK12.egSYMBOL))) ## get all genes involved in their regulatory modules from ## the RegulonDB data in this package mt <- match(filtered.regulon6.1[,"EgID_TF"], TFgenesEgIDs) allGenes <- as.character(unique(as.vector( as.matrix(filtered.regulon6.1[!is.na(mt), c("EgID_TF","EgID_TG")])))) mtTF <- match(filtered.regulon6.1[,"EgID_TF"],allGenes) mtTG <- match(filtered.regulon6.1[,"EgID_TG"],allGenes) ## select the corresponding subset of the RegulonDB data in this package subset.filtered.regulon6.1 <- filtered.regulon6.1[!is.na(mtTF) & !is.na(mtTG),] TFi <- match(subset.filtered.regulon6.1[,"EgID_TF"], allGenes) TGi <- match(subset.filtered.regulon6.1[,"EgID_TG"], allGenes) subset.filtered.regulon6.1 <- cbind(subset.filtered.regulon6.1, idx_TF=TFi, idx_TG=TGi) ## build an adjacency matrix representing the transcriptional regulatory ## relationships from these regulatory modules p <- length(allGenes) adjacencyMatrix <- matrix(FALSE, nrow=p, ncol=p) rownames(adjacencyMatrix) <- colnames(adjacencyMatrix) <- allGenes idxTFTG <- as.matrix(subset.filtered.regulon6.1[,c("idx_TF","idx_TG")]) adjacencyMatrix[idxTFTG] <- adjacencyMatrix[cbind(idxTFTG[,2],idxTFTG[,1])] <- TRUE ## calculate functional coherence on these regulatory modules fc <- qpFunctionalCoherence(adjacencyMatrix, TFgenes=TFgenesEgIDs, chip="org.EcK12.eg.db") print(sprintf("the %s module has a FC value of %.2f", mget(names(fc$functionalCoherenceValues),org.EcK12.egSYMBOL), fc$functionalCoherenceValues)) ## End(Not run)
Builds a positive definite matrix from an undirected graph G that can be used as a covariance matrix for a Gaussian graphical model with graph G. The inverse of the resulting matrix contains zeroes at the missing edges of the given undirected graph G.
qpG2Sigma(g, rho=0, matrix.completion=c("HTF", "IPF"), tol=0.001, verbose=FALSE, R.code.only=FALSE)qpG2Sigma(g, rho=0, matrix.completion=c("HTF", "IPF"), tol=0.001, verbose=FALSE, R.code.only=FALSE)
g |
undirected graph specified either as a |
rho |
real number between -1/(n.var-1) and 1 corresponding to the mean marginal correlation |
matrix.completion |
algorithm to employ in the matrix completion operations
employed to construct a positive definite matrix with the
zero pattern specified in |
tol |
tolerance under which the matrix completion algorithm stops. |
verbose |
show progress on the calculations. |
R.code.only |
logical; if FALSE then the faster C implementation is used in the internal call to the HTF, or IPF, algorithm (default); if TRUE then only R code is executed. |
The random covariance matrix is built by first generating a random matrix
with the function qpRndWishart from a Wishart distribution
whose expected value is a matrix with unit diagonal and constant off-diagonal
entries equal to rho.
A random positive definite matrix that can be used as a covariance matrix
for a Gaussian graphical model with graph G.
A. Roverato
Tur, I., Roverato, A. and Castelo, R. Mapping eQTL networks with mixed graphical Markov models. Genetics, 198(4):1377-1393, 2014.
qpRndGraph
qpGetCliques
qpIPF
qpRndWishart
rmvnorm
set.seed(123) G <- qpRndGraph(p=5, d=2) Sigma <- qpG2Sigma(G, rho=0.5) round(solve(Sigma), digits=2) as(G, "matrix")set.seed(123) G <- qpRndGraph(p=5, d=2) Sigma <- qpG2Sigma(G, rho=0.5) round(solve(Sigma), digits=2) as(G, "matrix")
Estimates generalized non-rejection rates for every pair of variables from two or more data sets.
## S4 method for signature 'ExpressionSet' qpGenNrr(X, datasetIdx=1, qOrders=NULL, I=NULL, restrict.Q=NULL, fix.Q=NULL, return.all=FALSE, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'data.frame' qpGenNrr(X, datasetIdx=1, qOrders=NULL, I=NULL, restrict.Q=NULL, fix.Q=NULL, return.all=FALSE, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'matrix' qpGenNrr(X, datasetIdx=1, qOrders=NULL, I=NULL, restrict.Q=NULL, fix.Q=NULL, return.all=FALSE, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10)## S4 method for signature 'ExpressionSet' qpGenNrr(X, datasetIdx=1, qOrders=NULL, I=NULL, restrict.Q=NULL, fix.Q=NULL, return.all=FALSE, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'data.frame' qpGenNrr(X, datasetIdx=1, qOrders=NULL, I=NULL, restrict.Q=NULL, fix.Q=NULL, return.all=FALSE, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'matrix' qpGenNrr(X, datasetIdx=1, qOrders=NULL, I=NULL, restrict.Q=NULL, fix.Q=NULL, return.all=FALSE, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10)
X |
data set from where to estimate the average non-rejection rates. It can be an ExpressionSet object, a data frame or a matrix. |
datasetIdx |
either a single number, or a character string, indicating
the column in the phenotypic data of the |
qOrders |
either a NULL value (default) indicating that a default guess on the q-order will be employed for each data set or a vector of particular orders with one for each data set. The default guess corresponds to the floor of the median value among the valid q orders of the data set. |
I |
indexes or names of the variables in |
restrict.Q |
indexes or names of the variables in |
fix.Q |
indexes or names of the variables in |
return.all |
logical; if TRUE all intervining non-rejection rates will be return in a matrix per dataset within a list; FALSE (default) if only generalized non-rejection rates should be returned. |
nTests |
number of tests to perform for each pair for variables. |
alpha |
significance level of each test. |
pairup.i |
subset of vertices to pair up with subset |
pairup.j |
subset of vertices to pair up with subset |
long.dim.are.variables |
logical; if TRUE it is assumed that when the data is a data frame or a matrix, the longer dimension is the one defining the random variables; if FALSE, then random variables are assumed to be at the columns of the data frame or matrix. |
verbose |
show progress on the calculations. |
identicalQs |
use identical conditioning subsets for every pair of vertices
(default), otherwise sample a new collection of |
exact.test |
logical; if |
use |
a character string defining the way in which calculations are done in the
presence of missing values. It can be either |
tol |
maximum tolerance controlling the convergence of the EM algorithm employed
when the argument |
R.code.only |
logical; if FALSE then the faster C implementation is used (default); if TRUE then only R code is executed. |
clusterSize |
size of the cluster of processors to employ if we wish to
speed-up the calculations by performing them in parallel. A value of 1
(default) implies a single-processor execution. The use of a cluster of
processors requires having previously loaded the packages |
estimateTime |
logical; if |
nAdj2estimateTime |
number of adjacencies to employ when estimating the
time of calculations ( |
Note that when specifying a vector of particular orders q, these values
should be in the range 1 to min(p,n-3), where p is the number of
variables and n the number of observations for the corresponding data set.
The computational cost increases linearly within each q value and
quadratically in p. When setting identicalQs to FALSE the
computational cost may increase between 2 times and one order of magnitude
(depending on p and q) while asymptotically the estimation of the
non-rejection rate converges to the same value.
When I is set different to NULL then mixed graphical model theory
is employed and, concretely, it is assumed that the data comes from an homogeneous
conditional Gaussian distribution. In this setting further restrictions to the
maximum value of q apply, concretely, it cannot be smaller than
p plus the number of levels of the discrete variables involved in the
marginal distributions employed by the algorithm. By default, with
exact.test=TRUE, an exact test for conditional independence is employed,
otherwise an asymptotic one will be used. Full details on these features can
be found in Tur, Roverato and Castelo (2014).
A list containing the following two or more entries: a first one with name
genNrr with a dspMatrix-class symmetric matrix of estimated
generalized non-rejection rates with the diagonal set to NA values. When
using the arguments pairup.i and pairup.j, those cells outside the
constraint pairs will get also a NA value; a second one with name
qOrders with the q-orders employed in the calculation for each data set;
if return.all=TRUE then there will be one additional entry for each data
set containing the matrix of the non-rejection rates estimated from that data
set with the corresponding q-order, using the indexing value of the data set as
entry name.
Note, however, that when estimateTime=TRUE, then instead of the list with
matrices of estimated (generalized) non-rejection rates, a vector specifying the
estimated number of days, hours, minutes and seconds for completion of the
calculations is returned.
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. Reverse engineering molecular regulatory networks from microarray data with qp-graphs. J. Comp. Biol., 16(2):213-227, 2009.
Tur, I., Roverato, A. and Castelo, R. Mapping eQTL networks with mixed graphical Markov models. Genetics, 198:1377-1393, 2014.
qpNrr
qpAvgNrr
qpEdgeNrr
qpHist
qpGraphDensity
qpClique
nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) ## simulate two independent Gaussian graphical models determined ## by two undirected d-regular graphs model1 <- rUGgmm(dRegularGraphParam(p=nVar, d=maxCon), rho=0.5) model2 <- rUGgmm(dRegularGraphParam(p=nVar, d=maxCon), rho=0.5) ## simulate two independent data sets from the previous graphical models X1 <- rmvnorm(nObs, model1) dim(X1) X2 <- rmvnorm(nObs, model2) dim(X2) ## estimate generalized non-rejection rates from the joint data nrr.estimates <- qpGenNrr(rbind(X1, X2), datasetIdx=rep(1:2, each=nObs), qOrders=c("1"=5, "2"=5), long.dim.are.variables=FALSE, verbose=FALSE) ## create adjacency matrices from the undirected graphs ## determining the two Gaussian graphical models A1 <- as(model1$g, "matrix") == 1 A2 <- as(model2$g, "matrix") == 1 ## distribution of generalized non-rejection rates for the common present edges summary(nrr.estimates$genNrr[upper.tri(nrr.estimates$genNrr) & A1 & A2]) ## distribution of generalized non-rejection rates for the present edges specific to A1 summary(nrr.estimates$genNrr[upper.tri(nrr.estimates$genNrr) & A1 & !A2]) ## distribution of generalized non-rejection rates for the present edges specific to A2 summary(nrr.estimates$genNrr[upper.tri(nrr.estimates$genNrr) & !A1 & A2]) ## distribution of generalized non-rejection rates for the common missing edges summary(nrr.estimates$genNrr[upper.tri(nrr.estimates$genNrr) & !A1 & !A2]) ## compare with the average non-rejection rate on the pooled data set avgnrr.estimates <- qpNrr(rbind(X1, X2), q=5, long.dim.are.variables=FALSE, verbose=FALSE) ## distribution of average non-rejection rates for the common present edges summary(avgnrr.estimates[upper.tri(avgnrr.estimates) & A1 & A2]) ## distribution of average non-rejection rates for the present edges specific to A1 summary(avgnrr.estimates[upper.tri(avgnrr.estimates) & A1 & !A2]) ## distribution of average non-rejection rates for the present edges specific to A2 summary(avgnrr.estimates[upper.tri(avgnrr.estimates) & !A1 & A2]) ## distribution of average non-rejection rates for the common missing edges summary(avgnrr.estimates[upper.tri(avgnrr.estimates) & !A1 & !A2])nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) ## simulate two independent Gaussian graphical models determined ## by two undirected d-regular graphs model1 <- rUGgmm(dRegularGraphParam(p=nVar, d=maxCon), rho=0.5) model2 <- rUGgmm(dRegularGraphParam(p=nVar, d=maxCon), rho=0.5) ## simulate two independent data sets from the previous graphical models X1 <- rmvnorm(nObs, model1) dim(X1) X2 <- rmvnorm(nObs, model2) dim(X2) ## estimate generalized non-rejection rates from the joint data nrr.estimates <- qpGenNrr(rbind(X1, X2), datasetIdx=rep(1:2, each=nObs), qOrders=c("1"=5, "2"=5), long.dim.are.variables=FALSE, verbose=FALSE) ## create adjacency matrices from the undirected graphs ## determining the two Gaussian graphical models A1 <- as(model1$g, "matrix") == 1 A2 <- as(model2$g, "matrix") == 1 ## distribution of generalized non-rejection rates for the common present edges summary(nrr.estimates$genNrr[upper.tri(nrr.estimates$genNrr) & A1 & A2]) ## distribution of generalized non-rejection rates for the present edges specific to A1 summary(nrr.estimates$genNrr[upper.tri(nrr.estimates$genNrr) & A1 & !A2]) ## distribution of generalized non-rejection rates for the present edges specific to A2 summary(nrr.estimates$genNrr[upper.tri(nrr.estimates$genNrr) & !A1 & A2]) ## distribution of generalized non-rejection rates for the common missing edges summary(nrr.estimates$genNrr[upper.tri(nrr.estimates$genNrr) & !A1 & !A2]) ## compare with the average non-rejection rate on the pooled data set avgnrr.estimates <- qpNrr(rbind(X1, X2), q=5, long.dim.are.variables=FALSE, verbose=FALSE) ## distribution of average non-rejection rates for the common present edges summary(avgnrr.estimates[upper.tri(avgnrr.estimates) & A1 & A2]) ## distribution of average non-rejection rates for the present edges specific to A1 summary(avgnrr.estimates[upper.tri(avgnrr.estimates) & A1 & !A2]) ## distribution of average non-rejection rates for the present edges specific to A2 summary(avgnrr.estimates[upper.tri(avgnrr.estimates) & !A1 & A2]) ## distribution of average non-rejection rates for the common missing edges summary(avgnrr.estimates[upper.tri(avgnrr.estimates) & !A1 & !A2])
Finds the set of (maximal) cliques of a given undirected graph.
qpGetCliques(g, clqspervtx=FALSE, verbose=TRUE)qpGetCliques(g, clqspervtx=FALSE, verbose=TRUE)
g |
either a |
clqspervtx |
logical; if TRUE then the resulting list returned by the function includes additionally p entries at the beginning (p=number of variables) each corresponding to a vertex in the graph and containing the indices of the cliques where that vertex belongs to; if FALSE these additional entries are not included (default). |
verbose |
show progress on calculations. |
To find the list of all (maximal) cliques in an undirected graph is an NP-hard problem which means that its computational cost is bounded by an exponential running time (Garey and Johnson, 1979). For this reason, this is an extremely time and memory consuming computation for large dense graphs. The current implementation uses C code from the GNU GPL Cliquer library by Niskanen and Ostergard (2003).
A list of maximal cliques. When clqspervtx=TRUE the first p entries
(p=number of variables) contain, each of them, the indices of the cliques where
that particular vertex belongs to.
R. Castelo
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n. J. Mach. Learn. Res., 7:2621-2650, 2006.
Garey, M.R. and Johnson D.S. Computers and intractability: a guide to the theory of NP-completeness. W.H. Freeman, San Francisco, 1979.
Niskanen, S. Ostergard, P. Cliquer User's Guide, Version 1.0. Communications Laboratory, Helsinki University of Technology, Espoo, Finland, Tech. Rep. T48, 2003. (http://users.tkk.fi/~pat/cliquer.html)
require(graph) set.seed(123) nVar <- 50 g1 <- randomEGraph(V=as.character(1:nVar), p=0.3) clqs1 <- qpGetCliques(g1, verbose=FALSE) length(clqs1) summary(sapply(clqs1, length)) g2 <- randomEGraph(V=as.character(1:nVar), p=0.7) clqs2 <- qpGetCliques(g2, verbose=FALSE) length(clqs2) clqs2 <- qpGetCliques(g2, verbose=FALSE) summary(sapply(clqs2, length))require(graph) set.seed(123) nVar <- 50 g1 <- randomEGraph(V=as.character(1:nVar), p=0.3) clqs1 <- qpGetCliques(g1, verbose=FALSE) length(clqs1) summary(sapply(clqs1, length)) g2 <- randomEGraph(V=as.character(1:nVar), p=0.7) clqs2 <- qpGetCliques(g2, verbose=FALSE) length(clqs2) clqs2 <- qpGetCliques(g2, verbose=FALSE) summary(sapply(clqs2, length))
The "qpGraph" class is the class to store and manipulate
q-order (partial) correlation graphs, or qp-graphs for short. See
Castelo and Roverato (2006, 2009) for a mathematical and statistical
definition of a qp-graph.
In earlier versions 1.x of the qpgraph package there
was a function called qpGraph() to obtain a qp-graph from a
matrix of non-rejection rates. This function, as it was written,
has been deprecated and replaced by this class and corresponding
constructor methods of the same name. The main difference with respect
to earlier 1.x versions is that the argument threshold is now
called epsilon, the argument return.type has been
removed and the current version returns an object of this class
qpGraph described in this manual page.
epsilon |
maximum cutoff value met by the edges present in the qp-graph. |
topPairs |
number of edges from the top of the ranking, defined by the
non-rejection rates in |
pairup.i |
subset of vertices to pair up with subset |
pairup.j |
subset of vertices to pair up with subset |
q |
q-order employed to derive the input matrix of non-rejection rates
|
n |
when the input matrix of non-rejection rates |
Objects can be created by calls of the form qpGraph(nrrMatrix, ...)
corresponding to constructor methods that take as input a matrix of
non-rejection rates, calculated with qpNrr.
p:number of vertices, in one-to-one correspondence with random variables.
q:order of the qp-graph, always smaller than p-2.
n:when the qp-graph has been estimated from data, this is the number of observations in the data set, which must be larger than q+2.
epsilon:maximum cutoff value of the non-rejection rate met by the edges that are present in the qp-graph.
g:undirected graph structure of the qp-graph stored as a graphBAM-class object.
qpGraph(nrrMatrix, ...)Constructor method where nrrMatrix is a matrix of non-rejection rates.
show(object)Method to display some bits of information about
the qp-graph stored in the input argument object.
R. Castelo
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n, J. Mach. Learn. Res., 7:2621-2650, 2006.
Castelo, R. and Roverato, A. Reverse engineering molecular regulatory networks from microarray data with qp-graphs. J. Comp. Biol., 16(2):213-227, 2009.
Calculates and plots the graph density as function of the non-rejection rate.
qpGraphDensity(nrrMatrix, threshold.lim=c(0,1), breaks=5, plot=TRUE, qpGraphDensityOutput=NULL, density.digits=0, titlegd="graph density as function of threshold")qpGraphDensity(nrrMatrix, threshold.lim=c(0,1), breaks=5, plot=TRUE, qpGraphDensityOutput=NULL, density.digits=0, titlegd="graph density as function of threshold")
nrrMatrix |
matrix of non-rejection rates. |
threshold.lim |
range of threshold values on the non-rejection rate. |
breaks |
either a number of threshold bins or a vector of threshold breakpoints. |
plot |
logical; if TRUE makes a plot of the result; if FALSE it does not. |
qpGraphDensityOutput |
output from a previous call to
|
density.digits |
number of digits in the reported graph densities. |
titlegd |
main title to be shown in the plot. |
The estimate of the sparseness of the resulting qp-graphs is calculated as one minus the area enclosed under the curve of graph densities.
A list with the graph density as function of threshold and an estimate of the sparseness of the resulting qp-graphs across the thresholds.
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n, J. Mach. Learn. Res., 7:2621-2650, 2006.
qpNrr
qpAvgNrr
qpEdgeNrr
qpClique
require(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## the higher the q the sparser the qp-graph nrr.estimates <- qpNrr(X, q=1, verbose=FALSE) qpGraphDensity(nrr.estimates, plot=FALSE)$sparseness nrr.estimates <- qpNrr(X, q=5, verbose=FALSE) qpGraphDensity(nrr.estimates, plot=FALSE)$sparsenessrequire(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## the higher the q the sparser the qp-graph nrr.estimates <- qpNrr(X, q=1, verbose=FALSE) qpGraphDensity(nrr.estimates, plot=FALSE)$sparseness nrr.estimates <- qpNrr(X, q=5, verbose=FALSE) qpGraphDensity(nrr.estimates, plot=FALSE)$sparseness
Plots the distribution of non-rejection rates.
qpHist(nrrMatrix, A=NULL, titlehist = "all estimated\nnon-rejection rates", freq=TRUE)qpHist(nrrMatrix, A=NULL, titlehist = "all estimated\nnon-rejection rates", freq=TRUE)
nrrMatrix |
matrix of non-rejection rates. |
A |
adjacency matrix of an undirected graph whose present and missing edges will be employed to show separately the distribution of non-rejection rates. |
titlehist |
main title of the histogram(s). |
freq |
logical; if TRUE, the histograms show frequencies (counts) of occurrence of the different non-rejection rate values; if FALSE, then probability densities are plotted |
This function plots histograms using the R-function hist and therefore
the way they are displayed follows that of this R-function.
None
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n, J. Mach. Learn. Res., 7:2621-2650, 2006.
qpNrr
qpAvgNrr
qpEdgeNrr
qpGraphDensity
qpClique
require(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) nrr.estimates <- qpNrr(X, q=5, verbose=FALSE) qpHist(nrr.estimates, A)require(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) nrr.estimates <- qpNrr(X, q=5, verbose=FALSE) qpHist(nrr.estimates, A)
Performs maximum likelihood estimation of a covariance matrix given the independence constraints from an input undirected graph.
qpHTF(S, g, tol = 0.001, verbose = FALSE, R.code.only = FALSE)qpHTF(S, g, tol = 0.001, verbose = FALSE, R.code.only = FALSE)
S |
input matrix, in the context of this package, the sample covariance matrix. |
g |
input undirected graph. |
tol |
tolerance under which the iterative algorithm stops. |
verbose |
show progress on calculations. |
R.code.only |
logical; if FALSE then the faster C implementation is used (default); if TRUE then only R code is executed. |
This is an alternative to the Iterative Proportional Fitting (IPF) algorithm
(see, Whittaker, 1990, pp. 182-185 and qpIPF) which also
adjusts the input matrix to the independence constraints in the input undirected
graph. However, differently to the IPF, it works by going through each of the
vertices fitting the marginal distribution over the corresponding vertex boundary.
It stops when the adjusted matrix at the current iteration differs from the matrix
at the previous iteration in less or equal than a given tolerance value. This
algorithm is described by Hastie, Tibshirani and Friedman (2009, pg. 634), hence we
name it here HTF, and it has the advantage over the IPF that it does not require the
list of maximal cliques of the graph which may be exponentially large. In
contrast, it requires that the maximum boundary size of the graph is below the
number of samples where the input sample covariance matrix S was estimated.
For the purpose of exploring qp-graphs that meet such a requirement, one can use
the function qpBoundary.
The input matrix adjusted to the constraints imposed by the input undirected graph, i.e., a maximum likelihood estimate of the sample covariance matrix that includes the independence constraints encoded in the undirected graph.
Thanks to Giovanni Marchetti for bringing us our attention to this algorithm and
sharing an early version of its implementation on the R package ggm.
R. Castelo
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n. J. Mach. Learn. Res., 7:2621-2650, 2006.
Hastie, T., Tibshirani, R. and Friedman, J.H. The Elements of Statistical Learning, Springer, 2009.
Tur, I., Roverato, A. and Castelo, R. Mapping eQTL networks with mixed graphical Markov models. Genetics, 198(4):1377-1393, 2014.
Whittaker, J. Graphical Models in Applied Multivariate Statistics. Wiley, 1990.
require(graph) require(mvtnorm) nVar <- 50 ## number of variables nObs <- 100 ## number of observations to simulate set.seed(123) g <- randomEGraph(as.character(1:nVar), p=0.15) Sigma <- qpG2Sigma(g, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## MLE of the sample covariance matrix S <- cov(X) ## more efficient MLE of the sample covariance matrix using HTF S_htf <- qpHTF(S, g) ## get the adjacency matrix and put the diagonal to one A <- as(g, "matrix") diag(A) <- 1 ## entries in S and S_htf for present edges in g should coincide max(abs(S_htf[A==1] - S[A==1])) ## entries in the inverse of S_htf for missing edges in g should be zero max(solve(S_htf)[A==0])require(graph) require(mvtnorm) nVar <- 50 ## number of variables nObs <- 100 ## number of observations to simulate set.seed(123) g <- randomEGraph(as.character(1:nVar), p=0.15) Sigma <- qpG2Sigma(g, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## MLE of the sample covariance matrix S <- cov(X) ## more efficient MLE of the sample covariance matrix using HTF S_htf <- qpHTF(S, g) ## get the adjacency matrix and put the diagonal to one A <- as(g, "matrix") diag(A) <- 1 ## entries in S and S_htf for present edges in g should coincide max(abs(S_htf[A==1] - S[A==1])) ## entries in the inverse of S_htf for missing edges in g should be zero max(solve(S_htf)[A==0])
Imports non-rejection rates from an external flat file.
qpImportNrr(filename, nTests)qpImportNrr(filename, nTests)
filename |
name of the flat file with the data on the non-rejection rates. |
nTests |
number of tests performed in the estimation of these non-rejection rates. |
This function expects a flat file with three tab-separated columns corresponding to, respectively, 0-based index of one of the variables, 0-based index of the other variable, number of non-rejected tests for the pair of variables of that row in the text file. An example of a few lines of that file would be:
6 3 95 6 4 98 6 5 23 7 0 94 7 1 94
After reading the file the function builds a matrix of non-rejection rates by
dividing the number of non-rejected tests by nTests. Note that if the flat
file to be imported would eventually have directly the rates instead of the
number of tests, these can be also imported by setting nTests=1.
This function is thought to be used to read files obtained from the standalone
parallel version of qpNrr which can be downloaded from
http://functionalgenomics.upf.edu/qp.
A symmetric matrix of non-rejection rates with the diagonal set to the NA
value.
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n, J. Mach. Learn. Res., 7:2621-2650, 2006.
Performs maximum likelihood estimation of a covariance matrix given the independence constraints from an input list of (maximal) cliques.
qpIPF(vv, clqlst, tol = 0.001, verbose = FALSE, R.code.only = FALSE)qpIPF(vv, clqlst, tol = 0.001, verbose = FALSE, R.code.only = FALSE)
vv |
input matrix, in the context of this package, the sample covariance matrix. |
clqlst |
list of maximal cliques obtained from an undirected graph
by using the function |
tol |
tolerance under which the iterative algorithm stops. |
verbose |
show progress on calculations. |
R.code.only |
logical; if FALSE then the faster C implementation is used (default); if TRUE then only R code is executed. |
The Iterative proportional fitting algorithm (see, Whittaker, 1990, pp. 182-185) adjusts the input matrix to the independence constraints in the undirected graph from where the input list of cliques belongs to, by going through each of the cliques fitting the marginal distribution over the clique for the fixed conditional distribution of the clique. It stops when the adjusted matrix at the current iteration differs from the matrix at the previous iteration in less or equal than a given tolerance value.
The input matrix adjusted to the constraints imposed by the list of cliques, i.e., a maximum likelihood estimate of the sample covariance matrix that includes the independence constraints encoded in the undirected graph formed by the given list of cliques.
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n. J. Mach. Learn. Res., 7:2621-2650, 2006.
Tur, I., Roverato, A. and Castelo, R. Mapping eQTL networks with mixed graphical Markov models. Genetics, 198(4):1377-1393, 2014.
Whittaker, J. Graphical models in applied multivariate statistics. Wiley, 1990.
require(graph) require(mvtnorm) nVar <- 50 ## number of variables nObs <- 100 ## number of observations to simulate set.seed(123) g <- randomEGraph(as.character(1:nVar), p=0.15) Sigma <- qpG2Sigma(g, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## MLE of the sample covariance matrix S <- cov(X) ## more efficient MLE of the sample covariance matrix using IPF clqs <- qpGetCliques(g, verbose=FALSE) S_ipf <- qpIPF(S, clqs) ## get the adjacency matrix and put the diagonal to one A <- as(g, "matrix") diag(A) <- 1 ## entries in S and S_ipf for present edges in g should coincide max(abs(S_ipf[A==1] - S[A==1])) ## entries in the inverse of S_ipf for missing edges in g should be zero max(solve(S_ipf)[A==0])require(graph) require(mvtnorm) nVar <- 50 ## number of variables nObs <- 100 ## number of observations to simulate set.seed(123) g <- randomEGraph(as.character(1:nVar), p=0.15) Sigma <- qpG2Sigma(g, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## MLE of the sample covariance matrix S <- cov(X) ## more efficient MLE of the sample covariance matrix using IPF clqs <- qpGetCliques(g, verbose=FALSE) S_ipf <- qpIPF(S, clqs) ## get the adjacency matrix and put the diagonal to one A <- as(g, "matrix") diag(A) <- 1 ## entries in S and S_ipf for present edges in g should coincide max(abs(S_ipf[A==1] - S[A==1])) ## entries in the inverse of S_ipf for missing edges in g should be zero max(solve(S_ipf)[A==0])
Obtains partial correlation coefficients from a given concentration matrix.
qpK2ParCor(K)qpK2ParCor(K)
K |
positive definite matrix, typically a concentration matrix. |
This function applies cov2cor to the given concentration
matrix and then changes the sign of the off-diagonal entries in order
to obtain a partial correlation matrix.
A partial correlation matrix.
R. Castelo and A. Roverato
Lauritzen, S.L. Graphical models. Oxford University Press, 1996.
require(graph) n.var <- 5 # number of variables set.seed(123) g <- randomEGraph(as.character(1:n.var), p=0.15) Sigma <- qpG2Sigma(g, rho=0.5) K <- solve(Sigma) round(qpK2ParCor(K), digits=2) as(g, "matrix")require(graph) n.var <- 5 # number of variables set.seed(123) g <- randomEGraph(as.character(1:n.var), p=0.15) Sigma <- qpG2Sigma(g, rho=0.5) K <- solve(Sigma) round(qpK2ParCor(K), digits=2) as(g, "matrix")
Estimates non-rejection rates for every pair of variables.
## S4 method for signature 'ExpressionSet' qpNrr(X, q=1, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'cross' qpNrr(X, q=1, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'data.frame' qpNrr(X, q=1, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'matrix' qpNrr(X, q=1, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10)## S4 method for signature 'ExpressionSet' qpNrr(X, q=1, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'cross' qpNrr(X, q=1, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'data.frame' qpNrr(X, q=1, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10) ## S4 method for signature 'matrix' qpNrr(X, q=1, I=NULL, restrict.Q=NULL, fix.Q=NULL, nTests=100, alpha=0.05, pairup.i=NULL, pairup.j=NULL, long.dim.are.variables=TRUE, verbose=TRUE, identicalQs=TRUE, exact.test=TRUE, use=c("complete.obs", "em"), tol=0.01, R.code.only=FALSE, clusterSize=1, estimateTime=FALSE, nAdj2estimateTime=10)
X |
data set from where to estimate the non-rejection rates.
It can be an |
q |
partial-correlation order to be employed. |
I |
indexes or names of the variables in |
restrict.Q |
indexes or names of the variables in |
fix.Q |
indexes or names of the variables in |
nTests |
number of tests to perform for each pair for variables. |
alpha |
significance level of each test. |
pairup.i |
subset of vertices to pair up with subset |
pairup.j |
subset of vertices to pair up with subset |
long.dim.are.variables |
logical; if |
verbose |
show progress on the calculations. |
identicalQs |
use identical conditioning subsets for every pair of vertices
(default), otherwise sample a new collection of |
exact.test |
logical; if |
use |
a character string defining the way in which calculations are done in the
presence of missing values. It can be either |
tol |
maximum tolerance controlling the convergence of the EM algorithm employed
when the argument |
R.code.only |
logical; if |
clusterSize |
size of the cluster of processors to employ if we wish to
speed-up the calculations by performing them in parallel. A value of 1
(default) implies a single-processor execution. The use of a cluster of
processors requires having previously loaded the packages |
estimateTime |
logical; if |
nAdj2estimateTime |
number of adjacencies to employ when estimating the
time of calculations ( |
Note that for pure continuous data the possible values of q should be in the
range 1 to min(p, n-3), where p is the number of variables and
n the number of observations. The computational cost increases linearly
with q and quadratically in p. When setting identicalQs
to FALSE the computational cost may increase between 2 times and one
order of magnitude (depending on p and q) while asymptotically
the estimation of the non-rejection rate converges to the same value. Full
details on the calculation of the non-rejection rate can be found in
Castelo and Roverato (2006).
When I is set different to NULL then mixed graphical model theory
is employed and, concretely, it is assumed that the data comes from an homogeneous
conditional Gaussian distribution. In this setting further restrictions to the
maximum value of q apply, concretely, it cannot be smaller than
p plus the number of levels of the discrete variables involved in the
marginal distributions employed by the algorithm. By default, with
exact.test=TRUE, an exact test for conditional independence is employed,
otherwise an asymptotic one will be used. Full details on these features can
be found in Tur, Roverato and Castelo (2014).
The argument I specifying what variables are discrete actually applies only
when X is a matrix object since in the other cases data types are specified
for each data columns or slot.
In the case that X is a qtl/cross object, the default NULL
values in arguments pairup.i and pairup.j actually imply pairing
all markers and phenotypes with numerical phenotypes only (including integer phenotypes).
Likewise, the default argument restrict.Q=NULL implies setting restrict.Q
to all numeric phenotypes. Setting these arguments to values other than NULL
allows the user to use those particular values being set.
A dspMatrix-class symmetric matrix of estimated non-rejection
rates with the diagonal set to NA values. If arguments pairup.i
and pairup.j are employed, those cells outside the constrained pairs
will get also a NA value.
Note, however, that when estimateTime=TRUE, then instead of the matrix
of estimated non-rejection rates, a vector specifying the estimated number of
days, hours, minutes and seconds for completion of the calculations is returned.
R. Castelo, A. Roverato and I. Tur
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n, J. Mach. Learn. Res., 7:2621-2650, 2006.
Tur, I., Roverato, A. and Castelo, R. Mapping eQTL networks with mixed graphical Markov models. Genetics, 198(4):1377-1393, 2014.
qpAvgNrr
qpEdgeNrr
qpHist
qpGraphDensity
qpClique
nVar <- 50 ## number of variables maxCon <- 3 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) ## simulate an undirected Gaussian graphical model ## determined by some random undirected d-regular graph model <- rUGgmm(dRegularGraphParam(p=nVar, d=maxCon), rho=0.5) model ## simulate data from this model X <- rmvnorm(nObs, model) dim(X) ## estimate non-rejection rates with q=3 nrr.estimates <- qpNrr(X, q=3, verbose=FALSE) ## create an adjacency matrix of the undirected graph ## determining the undirected Gaussian graphical model A <- as(model$g, "matrix") == 1 ## distribution of non-rejection rates for the present edges summary(nrr.estimates[upper.tri(nrr.estimates) & A]) ## distribution of non-rejection rates for the missing edges summary(nrr.estimates[upper.tri(nrr.estimates) & !A]) ## Not run: ## using R code only this would take much more time qpNrr(X, q=3, R.code.only=TRUE, estimateTime=TRUE) ## only for moderate and large numbers of variables the ## use of a cluster of processors speeds up the calculations library(snow) library(rlecuyer) nVar <- 500 maxCon <- 3 model <- rUGgmm(dRegularGraphParam(p=nVar, d=maxCon), rho=0.5) X <- rmvnorm(nObs, model) system.time(nrr.estimates <- qpNrr(X, q=10, verbose=TRUE)) system.time(nrr.estimates <- qpNrr(X, q=10, verbose=TRUE, clusterSize=4)) ## End(Not run)nVar <- 50 ## number of variables maxCon <- 3 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) ## simulate an undirected Gaussian graphical model ## determined by some random undirected d-regular graph model <- rUGgmm(dRegularGraphParam(p=nVar, d=maxCon), rho=0.5) model ## simulate data from this model X <- rmvnorm(nObs, model) dim(X) ## estimate non-rejection rates with q=3 nrr.estimates <- qpNrr(X, q=3, verbose=FALSE) ## create an adjacency matrix of the undirected graph ## determining the undirected Gaussian graphical model A <- as(model$g, "matrix") == 1 ## distribution of non-rejection rates for the present edges summary(nrr.estimates[upper.tri(nrr.estimates) & A]) ## distribution of non-rejection rates for the missing edges summary(nrr.estimates[upper.tri(nrr.estimates) & !A]) ## Not run: ## using R code only this would take much more time qpNrr(X, q=3, R.code.only=TRUE, estimateTime=TRUE) ## only for moderate and large numbers of variables the ## use of a cluster of processors speeds up the calculations library(snow) library(rlecuyer) nVar <- 500 maxCon <- 3 model <- rUGgmm(dRegularGraphParam(p=nVar, d=maxCon), rho=0.5) X <- rmvnorm(nObs, model) system.time(nrr.estimates <- qpNrr(X, q=10, verbose=TRUE)) system.time(nrr.estimates <- qpNrr(X, q=10, verbose=TRUE, clusterSize=4)) ## End(Not run)
Estimates partial correlation coefficients (PACs) for a Gaussian graphical model with undirected graph G and their corresponding p-values for the null hypothesis of zero-partial correlation.
## S4 method for signature 'ExpressionSet' qpPAC(X, g, return.K=FALSE, tol=0.001, matrix.completion=c("HTF", "IPF"), verbose=TRUE, R.code.only=FALSE) ## S4 method for signature 'data.frame' qpPAC(X, g, return.K=FALSE, long.dim.are.variables=TRUE, tol=0.001, matrix.completion=c("HTF", "IPF"), verbose=TRUE, R.code.only=FALSE) ## S4 method for signature 'matrix' qpPAC(X, g, return.K=FALSE, long.dim.are.variables=TRUE, tol=0.001, matrix.completion=c("HTF", "IPF"), verbose=TRUE, R.code.only=FALSE)## S4 method for signature 'ExpressionSet' qpPAC(X, g, return.K=FALSE, tol=0.001, matrix.completion=c("HTF", "IPF"), verbose=TRUE, R.code.only=FALSE) ## S4 method for signature 'data.frame' qpPAC(X, g, return.K=FALSE, long.dim.are.variables=TRUE, tol=0.001, matrix.completion=c("HTF", "IPF"), verbose=TRUE, R.code.only=FALSE) ## S4 method for signature 'matrix' qpPAC(X, g, return.K=FALSE, long.dim.are.variables=TRUE, tol=0.001, matrix.completion=c("HTF", "IPF"), verbose=TRUE, R.code.only=FALSE)
X |
data set from where to estimate the partial correlation coefficients. It can be an ExpressionSet object, a data frame or a matrix. |
g |
either a |
return.K |
logical; if TRUE this function also returns the concentration
matrix |
long.dim.are.variables |
logical; if TRUE it is assumed
that when |
tol |
maximum tolerance in the application of the IPF algorithm. |
matrix.completion |
algorithm to employ in the matrix completion operations
employed to construct a positive definite matrix with the
zero pattern specified in |
verbose |
show progress on the calculations. |
R.code.only |
logical; if FALSE then the faster C implementation is used (default); if TRUE then only R code is executed. |
In the context of maximum likelihood estimation (MLE) of PACs it is a necessary
condition for the existence of MLEs that the sample size n is larger
than the clique number w(G) of the graph G. If the sample size
n is larger than the maximum boundary of the input graph bd(G),
then the default matrix completion algorithm HTF by Hastie, Tibshirani and
Friedman (2009) can be used (see the function qpHTF() for details),
which has the avantage that is faster than IPF (see the function
qpIPF() for details).
The PAC estimation is done by first obtaining a MLE of the covariance matrix
using the qpIPF function and the p-values are calculated based on
the estimation of the standard errors (see Roverato and Whittaker, 1996) and
performing Wald tests based on the asymptotic chi-squared distribution.
A list with two matrices, one with the estimates of the PACs and the other with
their p-values. If return.K=TRUE then the MLE of the inverse covariance is
also returned as part of the list.
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n. J. Mach. Learn. Res., 7:2621-2650, 2006.
Castelo, R. and Roverato, A. Reverse engineering molecular regulatory networks from microarray data with qp-graphs. J. Comp. Biol., 16(2):213-227, 2009.
Hastie, T., Tibshirani, R. and Friedman, J.H. The Elements of Statistical Learning, Springer, 2009.
Roverato, A. and Whittaker, J. Standard errors for the parameters of graphical Gaussian models. Stat. Comput., 6:297-302, 1996.
qpGraph
qpCliqueNumber
qpClique
qpGetCliques
qpIPF
require(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) nrr.estimates <- qpNrr(X, verbose=FALSE) qpg <- qpGraph(nrr.estimates, epsilon=0.5) qpg$g pac.estimates <- qpPAC(X, g=qpg, verbose=FALSE) ## distribution absolute values of the estimated ## partial correlation coefficients of the present edges summary(abs(pac.estimates$R[upper.tri(pac.estimates$R) & A])) ## distribution absolute values of the estimated ## partial correlation coefficients of the missing edges summary(abs(pac.estimates$R[upper.tri(pac.estimates$R) & !A]))require(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) nrr.estimates <- qpNrr(X, verbose=FALSE) qpg <- qpGraph(nrr.estimates, epsilon=0.5) qpg$g pac.estimates <- qpPAC(X, g=qpg, verbose=FALSE) ## distribution absolute values of the estimated ## partial correlation coefficients of the present edges summary(abs(pac.estimates$R[upper.tri(pac.estimates$R) & A])) ## distribution absolute values of the estimated ## partial correlation coefficients of the missing edges summary(abs(pac.estimates$R[upper.tri(pac.estimates$R) & !A]))
Calculates the path weight for a path of an undirected graph.
## S4 method for signature 'matrix' qpPathWeight(X, path, Q=integer(0), M=integer(0), normalized=TRUE, R.code.only=TRUE)## S4 method for signature 'matrix' qpPathWeight(X, path, Q=integer(0), M=integer(0), normalized=TRUE, R.code.only=TRUE)
X |
covariance matrix. |
path |
character vector of consecutive vertex names defining a path in an undirected graph. |
Q |
indexes or names of the variables in |
M |
indexes or names of the variables in |
normalized |
logical; TRUE (default) when the calculated path weight should be normalized so that weights are comparable between paths with different endpoints, and false otherwise. |
R.code.only |
logical; if FALSE then the faster C implementation is used (not yet available); if TRUE then only R code is executed (default). |
Calculation of path weights. This implementation is still under development and will give only correct results with either population covariance matrices or sample covariance matrices estimated from data with n >> p. Consult (Roverato and Castelo, 2017) for further details.
The calculated path weight for the given path.
R. Castelo and A. Roverato
Roverato, A. and Castelo, R. The networked partial correlation and its application to the analysis of genetic interactions. J. R. Stat. Soc. Ser. C-Appl. Stat., 66:647-665, 2017. http://dx.doi.org/10.1111/rssc.12166
## example in Figure 1 from (Castelo and Roverato, 2017) ## undirected graph on 9 vertices edg <- matrix(c(1, 4, 2, 4, 3, 4, 4, 5, 5, 6, 5, 7, 8, 9), ncol=2, byrow=TRUE) ## create a corresponding synthetic precision matrix with ## partial correlation values set to -0.4 for all present edges K <- matrix(0, nrow=9, ncol=9, dimnames=list(1:9, 1:9)) K[edg] <- -0.4 K <- K + t(K) diag(K) <- 1 ## calculate the corresponding covariance matrix S <- solve(K) ## calculate networked partial correlations for all present ## edges npc <- sapply(1:nrow(edg), function(i) qpPathWeight(S, edg[i, ])) ## note that while all partial correlations are zero for missing ## edges and are equal to -0.4 for present edges, the corresponding ## networked partial correlations are also zero for missing edges ## but may be different between them for present edges, depending on ## the connections between the vertices cbind(edg, npc)## example in Figure 1 from (Castelo and Roverato, 2017) ## undirected graph on 9 vertices edg <- matrix(c(1, 4, 2, 4, 3, 4, 4, 5, 5, 6, 5, 7, 8, 9), ncol=2, byrow=TRUE) ## create a corresponding synthetic precision matrix with ## partial correlation values set to -0.4 for all present edges K <- matrix(0, nrow=9, ncol=9, dimnames=list(1:9, 1:9)) K[edg] <- -0.4 K <- K + t(K) diag(K) <- 1 ## calculate the corresponding covariance matrix S <- solve(K) ## calculate networked partial correlations for all present ## edges npc <- sapply(1:nrow(edg), function(i) qpPathWeight(S, edg[i, ])) ## note that while all partial correlations are zero for missing ## edges and are equal to -0.4 for present edges, the corresponding ## networked partial correlations are also zero for missing edges ## but may be different between them for present edges, depending on ## the connections between the vertices cbind(edg, npc)
Estimates Pearson correlation coefficients (PCCs) and their corresponding P-values between all pairs of variables from an input data set.
## S4 method for signature 'ExpressionSet' qpPCC(X) ## S4 method for signature 'data.frame' qpPCC(X, long.dim.are.variables=TRUE) ## S4 method for signature 'matrix' qpPCC(X, long.dim.are.variables=TRUE)## S4 method for signature 'ExpressionSet' qpPCC(X) ## S4 method for signature 'data.frame' qpPCC(X, long.dim.are.variables=TRUE) ## S4 method for signature 'matrix' qpPCC(X, long.dim.are.variables=TRUE)
X |
data set from where to estimate the Pearson correlation coefficients. It can be an ExpressionSet object, a data frame or a matrix. |
long.dim.are.variables |
logical; if TRUE it is assumed
that when |
The calculations made by this function are the same as the ones made for
a single pair of variables by the function cor.test but for
all the pairs of variables in the data set, with the exception of the treatment
of missing values, since only complete observations across all variables in
X are used.
A list with two matrices, one with the estimates of the PCCs and the other with their P-values.
R. Castelo and A. Roverato
require(graph) require(mvtnorm) nVar <- 50 ## number of variables nObs <- 10 ## number of observations to simulate set.seed(123) g <- randomEGraph(as.character(1:nVar), p=0.15) Sigma <- qpG2Sigma(g, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) pcc.estimates <- qpPCC(X) ## get the corresponding boolean adjacency matrix A <- as(g, "matrix") == 1 ## Pearson correlation coefficients of the present edges summary(abs(pcc.estimates$R[upper.tri(pcc.estimates$R) & A])) ## Pearson correlation coefficients of the missing edges summary(abs(pcc.estimates$R[upper.tri(pcc.estimates$R) & !A]))require(graph) require(mvtnorm) nVar <- 50 ## number of variables nObs <- 10 ## number of observations to simulate set.seed(123) g <- randomEGraph(as.character(1:nVar), p=0.15) Sigma <- qpG2Sigma(g, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) pcc.estimates <- qpPCC(X) ## get the corresponding boolean adjacency matrix A <- as(g, "matrix") == 1 ## Pearson correlation coefficients of the present edges summary(abs(pcc.estimates$R[upper.tri(pcc.estimates$R) & A])) ## Pearson correlation coefficients of the missing edges summary(abs(pcc.estimates$R[upper.tri(pcc.estimates$R) & !A]))
Plots a map of associated pairs defined by adjusted p-values
qpPlotMap(p.valueMatrix, markerPos, genePos, chrLen, p.value=0.05, adjust.method="holm", xlab="Ordered Markers", ylab="Ordered Genes", main="", ...)qpPlotMap(p.valueMatrix, markerPos, genePos, chrLen, p.value=0.05, adjust.method="holm", xlab="Ordered Markers", ylab="Ordered Genes", main="", ...)
p.valueMatrix |
squared symmetric matrix with raw p-values for all pairs. |
markerPos |
two-column matrix containing chromosome and position of each genetic marker. |
genePos |
two-column matrix containing chromosome and position of each gene. |
chrLen |
named vector with chromosome lengths. Vector names should correspond to chromosome names, which are displayed in the axes of the plot. This vector should be ordered following the same convention for chromosomes in arguments |
p.value |
adjusted p-value cutoff. |
adjust.method |
method employed to adjust the raw p-values. It is passed in a call to |
xlab |
label for the x-axis. |
ylab |
label for the y-axis. |
main |
main title of the plot, set to the empty string by default. |
... |
further arguments passed to the |
This function plots a map of present associations, typically between genetic markers and gene expression profiles (i.e., eQTL associations), according to the chromosomal locations of both the genetic markers and the genes. The input argument p.valueMatrix should contain the raw p-values of these associations. Present associations are selected by a cutoff given in the p.value argument applied to the adjusted p-values.
The input raw p-values can be obtained with the function qpAllCItests.
The selected present associations are invisibly returned.
R. Castelo
## generate uniformly random p-values for synthetic associations ## between m genetic markers and g genes into a symmetric matrix m <- 100 g <- 100 p <- m + g markerids <- paste0("m", 1:m) geneids <- paste0("g", 1:g) rndpvalues <- matrix(0, nrow=p, ncol=p, dimnames=list(c(markerids, geneids), c(markerids, geneids))) rndpvalues[1:m,(m+1):p] <- runif(m*g) ## put significant cis associations rndpvalues[cbind(1:m, (m+1):p)] <- rnorm(m, mean=1e-4, sd=1e-2)^2 ## put one hotspot locus with significant, but somehat weaker, trans associations hotspotmarker <- sample(1:m, size=1) rndpvalues[cbind(hotspotmarker, (m+1):p)] <- rnorm(g, mean=1e-2, sd=1e-2)^2 ## make matrix symmetric rndpvalues <- rndpvalues + t(rndpvalues) stopifnot(isSymmetric(rndpvalues)) rndpvalues[1:m, 1:m] <- rndpvalues[(m+1):p,(m+1):p] <- NA ## create chromosomal map chrlen <- c("chr1"=1000) posmarkers <- matrix(c(rep(1, m), seq(1, chrlen, length.out=m)), nrow=m) posgenes <- matrix(c(rep(1, g), seq(1, chrlen, length.out=g)), nrow=g) rownames(posmarkers) <- paste0("m", 1:m) rownames(posgenes) <- paste0("g", 1:g) qpPlotMap(rndpvalues, posmarkers, posgenes, chrlen, cex=3)## generate uniformly random p-values for synthetic associations ## between m genetic markers and g genes into a symmetric matrix m <- 100 g <- 100 p <- m + g markerids <- paste0("m", 1:m) geneids <- paste0("g", 1:g) rndpvalues <- matrix(0, nrow=p, ncol=p, dimnames=list(c(markerids, geneids), c(markerids, geneids))) rndpvalues[1:m,(m+1):p] <- runif(m*g) ## put significant cis associations rndpvalues[cbind(1:m, (m+1):p)] <- rnorm(m, mean=1e-4, sd=1e-2)^2 ## put one hotspot locus with significant, but somehat weaker, trans associations hotspotmarker <- sample(1:m, size=1) rndpvalues[cbind(hotspotmarker, (m+1):p)] <- rnorm(g, mean=1e-2, sd=1e-2)^2 ## make matrix symmetric rndpvalues <- rndpvalues + t(rndpvalues) stopifnot(isSymmetric(rndpvalues)) rndpvalues[1:m, 1:m] <- rndpvalues[(m+1):p,(m+1):p] <- NA ## create chromosomal map chrlen <- c("chr1"=1000) posmarkers <- matrix(c(rep(1, m), seq(1, chrlen, length.out=m)), nrow=m) posgenes <- matrix(c(rep(1, g), seq(1, chrlen, length.out=g)), nrow=g) rownames(posmarkers) <- paste0("m", 1:m) rownames(posgenes) <- paste0("g", 1:g) qpPlotMap(rndpvalues, posmarkers, posgenes, chrlen, cex=3)
Plots a graph using the Rgraphviz library
qpPlotNetwork(g, vertexSubset=graph::nodes(g), boundary=FALSE, minimumSizeConnComp=2, pairup.i=NULL, pairup.j=NULL, highlight=NULL, annotation=NULL, layout=c("twopi", "dot", "neato", "circo", "fdp"))qpPlotNetwork(g, vertexSubset=graph::nodes(g), boundary=FALSE, minimumSizeConnComp=2, pairup.i=NULL, pairup.j=NULL, highlight=NULL, annotation=NULL, layout=c("twopi", "dot", "neato", "circo", "fdp"))
g |
graph to plot provided as a |
vertexSubset |
subset of vertices that define the induced subgraph to be plotted. |
boundary |
flag set to |
minimumSizeConnComp |
minimum size of the connected components to be plotted. |
pairup.i |
subset of vertices to pair up with subset |
pairup.j |
subset of vertices to pair up with subset |
highlight |
subset of vertices to highlight by setting the color font to red. |
annotation |
name of an annotation package to transform gene identifiers into gene symbols when vertices correspond to genes. |
layout |
layout argument for the Rgraphviz library that plots the network. Possible values are |
This function acts as a wrapper for the functionality provided by the Rgraphviz
package to plot graphs in R. It should help to plot networks obtained with methods from
theqpgraph package.
The plotted graph is invisibly returned as a graphNEL-class object.
R. Castelo
## Not run: require(Rgraphviz) rndassociations <- qpUnifRndAssociation(10) g <- qpAnyGraph(abs(rndassociations), threshold=0.7, remove="below") qpPlotNetwork(g) ## this does not work at the moment and should be fixed ## End(Not run)## Not run: require(Rgraphviz) rndassociations <- qpUnifRndAssociation(10) g <- qpAnyGraph(abs(rndassociations), threshold=0.7, remove="below") qpPlotNetwork(g) ## this does not work at the moment and should be fixed ## End(Not run)
Calculates the precision-recall curve (see Fawcett, 2006) for a given measure of association between all pairs of variables in a matrix.
qpPrecisionRecall(measurementsMatrix, refGraph, decreasing=TRUE, pairup.i=NULL, pairup.j=NULL, recallSteps=seq(0, 1, by=0.1))qpPrecisionRecall(measurementsMatrix, refGraph, decreasing=TRUE, pairup.i=NULL, pairup.j=NULL, recallSteps=seq(0, 1, by=0.1))
measurementsMatrix |
matrix containing the measure of association between all pairs of variables. |
refGraph |
a reference graph from which to calculate the precision-recall
curve provided either as an adjacency matrix, a two-column matrix of edges,
a |
decreasing |
logical; if TRUE then the measurements are ordered in decreasing order; if FALSE then in increasing order. |
pairup.i |
subset of vertices to pair up with subset |
pairup.j |
subset of vertices to pair up with subset |
recallSteps |
steps of the recall on which to calculate precision. |
The measurementsMatrix should be symmetric and may have also contain
NA values which will not be taken into account. That is an alternative
way to restricting the variable pairs with the parameters pairup.i and
pairup.j.
A matrix where rows correspond to recall steps and columns correspond, respetively, to the actual recall, the precision, the number of true positives at that recall rate and the threshold score that yields that recall rate.
R. Castelo and A. Roverato
Fawcett, T. An introduction to ROC analysis. Pattern Recogn. Lett., 27:861-874, 2006.
qpPRscoreThreshold
qpGraph
qpAvgNrr
qpPCC
require(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## estimate non-rejection rates nrr.estimates <- qpNrr(X, q=5, verbose=FALSE) ## estimate Pearson correlation coefficients pcc.estimates <- qpPCC(X) ## calculate area under the precision-recall curve ## for both sets of estimated values of association nrr.prerec <- qpPrecisionRecall(nrr.estimates, refGraph=A, decreasing=FALSE, recallSteps=seq(0, 1, 0.1)) f <- approxfun(nrr.prerec[, c("Recall", "Precision")]) integrate(f, 0, 1)$value pcc.prerec <- qpPrecisionRecall(abs(pcc.estimates$R), refGraph=A, recallSteps=seq(0, 1, 0.1)) f <- approxfun(pcc.prerec[, c("Recall", "Precision")]) integrate(f, 0, 1)$valuerequire(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) ## estimate non-rejection rates nrr.estimates <- qpNrr(X, q=5, verbose=FALSE) ## estimate Pearson correlation coefficients pcc.estimates <- qpPCC(X) ## calculate area under the precision-recall curve ## for both sets of estimated values of association nrr.prerec <- qpPrecisionRecall(nrr.estimates, refGraph=A, decreasing=FALSE, recallSteps=seq(0, 1, 0.1)) f <- approxfun(nrr.prerec[, c("Recall", "Precision")]) integrate(f, 0, 1)$value pcc.prerec <- qpPrecisionRecall(abs(pcc.estimates$R), refGraph=A, recallSteps=seq(0, 1, 0.1)) f <- approxfun(pcc.prerec[, c("Recall", "Precision")]) integrate(f, 0, 1)$value
Calculates the score threshold at a given precision or recall level from a given precision-recall curve.
qpPRscoreThreshold(preRecFun, level, recall.level=TRUE, max.score=9999999)qpPRscoreThreshold(preRecFun, level, recall.level=TRUE, max.score=9999999)
preRecFun |
precision-recall function (output from
|
level |
recall or precision level. |
recall.level |
logical; if TRUE then it is assumed that the value given in the level parameter corresponds to a desired level of recall; if FALSE then it is assumed a desired level of precision. |
max.score |
maximum score given by the method that produced the precision-recall function to an association. |
The score threshold at which a given level of precision or recall is attained by
the given precision-recall function. For levels that do not form part of the
given function their score is calculated by linear interpolation and for this
reason is important to carefully specify a proper value for the max.score
parameter.
R. Castelo and A. Roverato
Fawcett, T. An introduction to ROC analysis. Pattern Recogn. Lett., 27:861-874, 2006.
require(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) nrr.estimates <- qpNrr(X, q=1, verbose=FALSE) nrr.prerec <- qpPrecisionRecall(nrr.estimates, A, decreasing=FALSE, recallSteps=seq(0, 1, by=0.1)) qpPRscoreThreshold(nrr.prerec, level=0.5, recall.level=TRUE, max.score=0) qpPRscoreThreshold(nrr.prerec, level=0.5, recall.level=FALSE, max.score=0)require(mvtnorm) nVar <- 50 ## number of variables maxCon <- 5 ## maximum connectivity per variable nObs <- 30 ## number of observations to simulate set.seed(123) A <- qpRndGraph(p=nVar, d=maxCon) Sigma <- qpG2Sigma(A, rho=0.5) X <- rmvnorm(nObs, sigma=as.matrix(Sigma)) nrr.estimates <- qpNrr(X, q=1, verbose=FALSE) nrr.prerec <- qpPrecisionRecall(nrr.estimates, A, decreasing=FALSE, recallSteps=seq(0, 1, by=0.1)) qpPRscoreThreshold(nrr.prerec, level=0.5, recall.level=TRUE, max.score=0) qpPRscoreThreshold(nrr.prerec, level=0.5, recall.level=FALSE, max.score=0)
Samples an undirected d-regular graph approximately uniformly at random.
qpRndGraph(p=6, d=2, labels=1:p, exclude=NULL, verbose=FALSE, return.type=c("adjacency.matrix", "edge.list", "graphBAM", "graphNEL"), R.code.only=FALSE)qpRndGraph(p=6, d=2, labels=1:p, exclude=NULL, verbose=FALSE, return.type=c("adjacency.matrix", "edge.list", "graphBAM", "graphNEL"), R.code.only=FALSE)
p |
number of vertices. |
d |
degree of every vertex. |
labels |
vertex labels. |
exclude |
vector of vertices inducing edges that should be excluded from the sampled d-regular graph. |
verbose |
show progress on the calculations. |
return.type |
class of object to be returned by the function |
R.code.only |
logical; if |
This function implements the algorithm from Steger and Wormald (1999) for sampling undirected d-regular graphs from a probability distribution of all d-regular graphs on p vertices which is approximately uniform. More concretely, for all vertex degree values d that grow as a small power of p, all d-regular graphs on p vertices will have in the limit the same probability as p grows large. Steger and Wormald (1999, pg. 396) believe that for d >> sqrt(p) the resulting probability distribution will no longer be approximately uniform.
This function is provided in order to generate a random undirected graph
as input to the function qpG2Sigma which samples a random
covariance matrix whose inverse (aka, precision matrix) has zeroes on those
cells corresponding to the missing edges in the input graph. d-regular
graphs are useful for working with synthetic graphical models for two
reasons: one is that d-regular graph density is a linear function of d and
the other is that the minimum connectivity degree of two disconnected vertices
is an upper bound of their outer connectivity (see Castelo and Roverato,
2006, pg. 2646).
The adjacency matrix of the resulting graph.
R. Castelo and A. Roverato
Castelo, R. and Roverato, A. A robust procedure for Gaussian graphical model search from microarray data with p larger than n, J. Mach. Learn. Res., 7:2621-2650, 2006.
Steger, A. and Wormald, N.C. Generating random regular graphs quickly, Combinatorics, Probab. and Comput., 8:377-396.
set.seed(123) A <- qpRndGraph(p=50, d=3) summary(apply(A, 1, sum))set.seed(123) A <- qpRndGraph(p=50, d=3) summary(apply(A, 1, sum))
Random generation for the (n.var * n.var) Wishart distribution (see
Press, 1972) with matrix parameter A=diag(delta)%*%P%*%diag(delta) and
degrees of freedom df.
qpRndWishart(delta=1, P=0, df=NULL, n.var=NULL)qpRndWishart(delta=1, P=0, df=NULL, n.var=NULL)
delta |
a numeric vector of |
P |
a ( |
df |
degrees of freedom. |
n.var |
dimension of the Wishart matrix. It is required only when both delata and P are scalar. |
The degrees of freedom are df > n.var-1 and the expected value of the
distribution is equal to df * A. The random generator is based on the
algorithm of Odell and Feiveson (1966).
A list of two n.var * n.var matrices rW and meanW where
rW is a random value from the Wishart and meanW is the expected
value of the distribution.
A. Roverato
Odell, P.L. and Feiveson, A.G. A numerical procedure to generate a sample covariance matrix. J. Am. Statist. Assoc. 61, 199-203, 1966.
Press, S.J. Applied Multivariate Analysis: Using Bayesian and Frequentist Methods of Inference. New York: Holt, Rinehalt and Winston, 1972.
Tur, I., Roverato, A. and Castelo, R. Mapping eQTL networks with mixed graphical Markov models. Genetics, 198(4):1377-1393, 2014.
## Construct an adjacency matrix for a graph on 6 vertices nVar <- 6 A <- matrix(0, nVar, nVar) A[1,2] <- A[2,3] <- A[3,4] <- A[3,5] <- A[4,6] <- A[5,6] <- 1 A=A + t(A) A set.seed(123) M <- qpRndWishart(delta=sqrt(1/nVar), P=0.5, n.var=nVar) M set.seed(123) d=1:6 M <- qpRndWishart(delta=d, P=0.7, df=20) M## Construct an adjacency matrix for a graph on 6 vertices nVar <- 6 A <- matrix(0, nVar, nVar) A[1,2] <- A[2,3] <- A[3,4] <- A[3,5] <- A[4,6] <- A[5,6] <- 1 A=A + t(A) A set.seed(123) M <- qpRndWishart(delta=sqrt(1/nVar), P=0.5, n.var=nVar) M set.seed(123) d=1:6 M <- qpRndWishart(delta=d, P=0.7, df=20) M
Report a top number of pairs of variables according to either some association measure and/or occurring in a given reference graph.
qpTopPairs(measurementsMatrix=NULL, refGraph=NULL, n=6L, file=NULL, decreasing=FALSE, pairup.i=NULL, pairup.j=NULL, annotation=NULL, fcOutput=NULL, fcOutput.na.rm=FALSE, digits=NULL)qpTopPairs(measurementsMatrix=NULL, refGraph=NULL, n=6L, file=NULL, decreasing=FALSE, pairup.i=NULL, pairup.j=NULL, annotation=NULL, fcOutput=NULL, fcOutput.na.rm=FALSE, digits=NULL)
measurementsMatrix |
matrix containing the measure of association between all pairs of variables. |
refGraph |
a reference graph containing the pairs that should be reported
and provided either as an adjacency matrix, a |
n |
number of pairs to report, 6 by default, use |
file |
file name to dump the pairs information as tab-separated column text. |
decreasing |
logical; if TRUE then the measurements are employed to be ordered in decreasing order; if FALSE then in increasing order. |
pairup.i |
subset of vertices to pair up with subset |
pairup.j |
subset of vertices to pair up with subset |
annotation |
name of an annotation package to transform gene identifiers into gene symbols when variables correspond to genes. |
fcOutput |
output of |
fcOutput.na.rm |
flag set to TRUE when pairs with |
digits |
number of decimal digits reported in the values of |
The measurementsMatrix should be symmetric and may have also contain
NA values which will not be taken into account. That is an alternative
way to restricting the variable pairs with the parameters pairup.i and
pairup.j. The same holds for refGraph. One of these two, should
be specified.
The ranking of pairs is invisibly returned.
R. Castelo
qpGraph
qpPrecisionRecall
qpFunctionalCoherence
qpTopPairs(matrix(runif(100), nrow=10, dimnames=list(1:10,1:10)))qpTopPairs(matrix(runif(100), nrow=10, dimnames=list(1:10,1:10)))
Builds a matrix of uniformly random association values between -1 and +1 for all pairs of variables that follow from the number of variables given as input argument.
qpUnifRndAssociation(n.var, var.names=as.character(1:n.var))qpUnifRndAssociation(n.var, var.names=as.character(1:n.var))
n.var |
number of variables. |
var.names |
names of the variables to use as row and column names in the resulting matrix. |
This function simply generates uniformly random association values with no independence
pattern associated to them. For generating a random covariance matrix that reflects
such a pattern use the function qpG2Sigma.
A symmetric matrix of uniformly random association values between -1 and +1.
R. Castelo
rndassociation <- qpUnifRndAssociation(100) summary(rndassociation[upper.tri(rndassociation)])rndassociation <- qpUnifRndAssociation(100) summary(rndassociation[upper.tri(rndassociation)])
Updates the set of (maximal) cliques of a given undirected graph when removing one edge.
qpUpdateCliquesRemoving(g, clqlst, v, w, verbose=TRUE)qpUpdateCliquesRemoving(g, clqlst, v, w, verbose=TRUE)
g |
either a |
clqlst |
list of cliques of the graph encoded in g. this list should
start on element n+1 (for n vertices) while between elements
1 to n there should be references to the cliques to which each
of the 1 to n vertices belong to (i.e., the output of
|
v |
vertex of the edge being removed. |
w |
vertex of the edge being removed. |
verbose |
show progress on calculations. |
To find the list of all (maximal) cliques in an undirected graph is an NP-hard
problem which means that its computational cost is bounded by an exponential
running time (Garey and Johnson, 1979). For this reason, this is an extremely
time and memory consuming computation for large dense graphs. If we spend the time
to obtain one such list of cliques and we remove one edge of the graph with this
function we may be able to update the set of maximal cliques instead of having
to generate it again entirely with qpGetCliques but it requires that
in the first call to qpGetCliques we set clqspervtx=TRUE.
It calls a C implementation of the algorithm from Stix (2004).
The updated list of maximal cliques after removing one edge from the input graph.
Note that because the corresponding input clique list had to be generated with the
argument clqspervtx=TRUE in the call to qpGetCliques, the
resulting updated list of cliques also includes in its first p entries
(p=number of variables) the indices of the cliques where that particular vertex
belongs to. Notice also that although this strategy might be in general more
efficient than generating again the entire list of cliques, when removing one edge
from the graph, the clique enumeration problem remains NP-hard (see Garey and Johnson, 1979)
and therefore depending on the input graph its computation may become unfeasible.
R. Castelo
Garey, M.R. and Johnson D.S. Computers and intractability: a guide to the theory of NP-completeness. W.H. Freeman, San Francisco, 1979.
Stix, V. Finding all maximal cliques in dynamic graphs Comput. Optimization and Appl., 27:173-186, 2004.
qpCliqueNumber
qpGetCliques
qpIPF
## the example below takes about 30 seconds to execute and for that reason ## it is not executed by default ## Not run: require(graph) set.seed(123) nVar <- 1000 g1 <- randomEGraph(V=as.character(1:nVar), p=0.1) g1 clqs1 <- qpGetCliques(g1, clqspervtx=TRUE, verbose=FALSE) length(clqs1) g2 <- removeEdge(from="1", to=edges(g1)[["1"]][1], g1) g2 system.time(clqs2a <- qpGetCliques(g2, verbose=FALSE)) system.time(clqs2b <- qpUpdateCliquesRemoving(g1, clqs1, "1", edges(g1)[["1"]][1], verbose=FALSE)) length(clqs2a) length(clqs2b)-nVar ## End(Not run)## the example below takes about 30 seconds to execute and for that reason ## it is not executed by default ## Not run: require(graph) set.seed(123) nVar <- 1000 g1 <- randomEGraph(V=as.character(1:nVar), p=0.1) g1 clqs1 <- qpGetCliques(g1, clqspervtx=TRUE, verbose=FALSE) length(clqs1) g2 <- removeEdge(from="1", to=edges(g1)[["1"]][1], g1) g2 system.time(clqs2a <- qpGetCliques(g2, verbose=FALSE)) system.time(clqs2b <- qpUpdateCliquesRemoving(g1, clqs1, "1", edges(g1)[["1"]][1], verbose=FALSE)) length(clqs2a) length(clqs2b)-nVar ## End(Not run)
The "SsdMatrix" class is the class of symmetric, dense matrices
in packed storage (just as a dspMatrix-class, i.e.,
only the upper triangle is stored) defined within the qpgraph
package to store corrected, or uncorrected, matrices of the sum of squares
and deviations (SSD) of pairs of random variables. A corrected SSD matrix
corresponds to a sample covariance matrix.
Objects can be created by calls of the form new("SsdMatrix", ...)
or by using qpCov() which estimates a sample covariance
matrix from data returning an object of this class.
ssd:Object of class dspMatrix-class
storing the SSD matrix.
n:Object of class "numeric" storing the sample
size employed to estimate the SSD matrix stored in the slot ssd.
This is specially relevant when the SSD matrix was estimated from data
with missing values by using complete observations only, which is the
default mode of operation of qpCov().
"SsdMatrix" extends class "dspMatrix", directly.
signature(x = "SsdMatrix")
signature(x = "SsdMatrix")
signature(object = "SsdMatrix")
signature(object = "SsdMatrix")
signature(object = "SsdMatrix", logarithm = "missing")
signature(object = "SsdMatrix", logarithm = "logical")
The "UGgmm" class is the class of undirected Gaussian graphical
Markov models defined within the qpgraph package to store
simulate and manipulate this type of graphical Markov models (GMMs).
An undirected Gaussian GMM is a family of multivariate normal distributions sharing a set of conditional independences encoded by means of an undirected graph. Further details can be found in the book of Lauritzen (1996).
Objects can be created by calls of the form UGgmm(g, ...) corresponding
to constructor methods or rUGgmm(n, g, ...) corresponding to random
simulation methods.
p:Object of class "integer" storing the dimension of the
undirected Gaussian GMM corresponding to the number of random variables.
g:Object of class graphBAM-class storing
the associated undirected labeled graph.
mean:Object of class "numeric" storing the mean vector.
sigma:Object of class dspMatrix-class
storing the covariance matrix.
UGgmm(g)Constructor method where g can be either an
adjacency matrix or a graphBAM-class object.
rUGgmm(n, g)Constructor simulation method that allows one to
simulate undirected Gaussian GMMs where n is the number of GMMs to
simulate and g can be either a graphParam object,
an adjacency matrix or a graphBAM-class object.
names(x)Accessor method to obtain the names of the
elements in the object x that can be retrieved with the $
accessor operator.
$Accessor operator to retrieve elements of the object
in an analogous way to a list.
dim(x)Dimension of the undirected Gaussian GMM corresponding to the total number of random variables.
dimnames(x)Names of the random variables in the undirected Gaussian GMM.
show(object)Method to display some bits of information about
the input undirected Gaussian GMM specified in object.
summary(object)Method to display a sumamry of the main features
of the input undirected Gaussian GMM specified in object.
plot(x, ...)Method to plot the undirected graph associated to the
the input undirected Gaussian GMM specified in x. It uses the plotting
capabilities from the Rgraphviz library to which further arguments
specified in ... are further passed.
R. Castelo
Lauritzen, S.L. Graphical models. Oxford University Press, 1996.