edd: new expression density diagnostics interface

In edd: expression density diagnostics

Description

this will replace edd.unsupervised; has more sensible parameters

Usage

edd(eset, distList=eddDistList, tx=c(sort,flatQQNormY)[[1]],
	refDist=c("multiSim", "theoretical")[1], 
	method=c("knn", "nnet", "test")[1], nRowPerCand=100, ...)

Arguments

eset	eset – instance of Biobase ExpressionSet class
distList	distList – list comprised of eddDist objects
tx	tx – transformation of data and reference prior to classification
refDist	refDist – type of reference distribution system to use
method	method – type of classifier to use. knn is k-nearest neighbors, nnet is neural net, test is max p-value from ks.test
nRowPerCand	nRowPerCand – number of realizations for a multiSim reference system
...	... – parameters to classifiers

Details

Classifies genes according to distributional shape, by comparing observed expression distributions to a collection of references, which may be simulated or evaluated theoretically.

The distList argument is important. It enumerates the catalog of distributions for classification of gene expression vectors by distributional shape. See the HOWTO-edd vignette for information on how this list is constructed and how it can be extended.

The tx argument specifies how the data are processed for comparison to the reference catalog. This is a function on a vector returning a vector, but the input and the output need not have the same length. The default value of tx is sort, which entails that the order statistics are treated as multivariate data for classification.

The refDist argument selects the type of reference catalog. Options are 'multiSim', for which the reference consists of nRowPerCand realizations of each catalog entry, and 'theoretical', for which the reference consists of one vector of quantiles for each catalog entry.

The method argument selects the type of classifier. It would be desirable to allow this to be a function, but there is insufficient structure on classifier argument and return value structure to permit this at present; see the e1071 package for some work on handling various classifiers programmatically (e.g., tune).

Value

a character vector or factor depending on the classifier

Author(s)

Vince Carey <[e-mail]>

See Also

ExpressionSet

Examples

require(Biobase)
data(sample.ExpressionSet)
# should filter to genes with reasonable variation
table( edd(sample.ExpressionSet, meth="nnet", size=10, decay=.2) )
library(golubEsets)
data(Golub_Merge)
madvec <- apply(exprs(Golub_Merge),1,mad)
minvec <- apply(exprs(Golub_Merge),1,min)
keep <- (madvec > median(madvec)) & (minvec > 300)
gmfilt <- Golub_Merge[keep==TRUE,]
ALL <- gmfilt$ALL.AML=="ALL"
gall <- gmfilt[,ALL==TRUE]
gaml <- gmfilt[,ALL==FALSE]
alldists <- edd(gall, meth="nnet", size=10, decay=.2)
amldists <- edd(gaml, meth="nnet", size=10, decay=.2)
table(alldists,amldists)
amldists2 <- edd(gaml, meth="nnet", refDist="theoretical", size=10, decay=.2)
table(amldists,amldists2)

edd documentation built on May 2, 2019, 6:10 p.m.