fit.rrfe: Reweighted Recursive Feature Elimination (RRFE)

In pathClass: Classification using biological pathways as prior knowledge

Description

Implementation of the Reweighted Recursive Feature Elimination (RRFE) algorithm. mapping must be a data.frame with at least two columns. The column names have to be c('probesetID','graphID'). Where 'probesetID' is the probeset ID present in the expression matrix (i.e. colnames(x)) and 'graphID' is any ID that represents the nodes in the graph (i.e. colnames(Gsub) or rownames(Gsub)). The purpose of the this mapping is that a gene or protein in the network might be represented by more than one probe set on the chip. Therefore, the algorithm must know which genes/protein in the network belongs to which probeset on the chip. However, the method is able to use all feature when one sets the parameter useAllFeatures to TRUE. When doing so, RRFE assigns the minimal wheight returned by GeneRank to those genes which are not present in Gsub.

Usage

  fit.rrfe(x, y, DEBUG = FALSE,
    scale = c("center", "scale"), Cs = 10^c(-3:3),
    stepsize = 0.1, useAllFeatures = F, mapping, Gsub,
    d = 0.5)

Arguments

x	a p x n matrix of expression measurements with p samples and n genes.
y	a factor of length p comprising the class labels.
DEBUG	should debugging information be plotted.
scale	a character vector defining if the data should be centered and/or scaled. Possible values are center and/or scale. Defaults to c('center', 'scale').
Cs	soft-margin tuning parameter of the SVM. Defaults to 10^c(-3:3).
stepsize	amount of features that are discarded in each step of the feature elimination. Defaults to 10%.
useAllFeatures	should all features be used for classification (see also Details).
mapping	a mapping that defines how probe sets are summarized to genes.
Gsub	an adjacency matrix that represents the underlying biological network.
d	the damping factor which controls the influence of the network data and the fold change on the ranking of the genes. Defaults to 0.5

Value

a RRFE fit object.

features	the selected features
error.bound	the span bound of the model
fit	the fitted SVM model

Note

The optimal number of features is found by using the span estimate. See Chapelle, O., Vapnik, V., Bousquet, O., and Mukherjee, S. (2002). Choosing multiple parameters for support vector machines. Machine Learning, 46(1), 131-159.

Author(s)

Marc Johannes [e-mail]

References

Johannes M, et al. (2010). Integration Of Pathway Knowledge Into A Reweighted Recursive Feature Elimination Approach For Risk Stratification Of Cancer Patients. Bioinformatics

Examples

## Not run: 
library(Biobase)
data(sample.ExpressionSet)
x <- t(exprs(sample.ExpressionSet))
y <- factor(pData(sample.ExpressionSet)$sex)
# create the mapping
library('hgu95av2.db')
mapped.probes <- mappedkeys(hgu95av2REFSEQ)
refseq <- as.list(hgu95av2REFSEQ[mapped.probes])
times <- sapply(refseq, length)
mapping <- data.frame(probesetID=rep(names(refseq), times=times), graphID=unlist(refseq), 
row.names=NULL, stringsAsFactors=FALSE)
mapping <- unique(mapping)
library(pathClass)
data(adjacency.matrix)
res.rrfe <- crossval(x, y, DEBUG=TRUE, theta.fit=fit.rrfe, folds=3, repeats=1, parallel=TRUE,
 Cs=10^(-3:3), mapping=mapping, Gsub=adjacency.matrix, d=1/2)

## End(Not run)

pathClass documentation built on May 29, 2017, 11:44 p.m.