meta.pca: META-ANALYTIC PRINCIPAL COMPONENT ANALYSIS IN INTEGRATIVE...

In metaOmic/metaPCA: META-ANALYTIC PRINCIPAL COMPONENT ANALYSIS IN INTEGRATIVE OMICS APPLICATION

Usage

meta.pca(DList, method, Meta.Dim, is.auto.Dim = TRUE, is.equal.Dim = FALSE, e.Dim, is.weight = TRUE, .var.quantile = 0.8, .scaleAdjust = FALSE, is.sparse = FALSE)

Arguments

DList	Input data set matrix (a list of multiple datasets; row=features, column=samples).
method	"SSC" = Sum of Squared Cosine, "SV" = Sum of variance
Meta.Dim	Dimension size of meta-eigenvector matrix
is.auto.Dim	Logical value whether dimension size of each study's eigenvector matrix (SSC) is determined by an arbitrary variance quantile
is.equal.Dim	Logical value whether dimension size of each study's eigenvector matrix (SSC) is equal across studies
e.Dim	Dimension size of each study's eigenvector matrix (SSC) when is.equal.Dim = TRUE
is.weight	Logical value whether the reciprocal of the largest eigenvalue is mutiplied to covariance matrix
.var.quantile	A threshold indicating the minimum variance of individual study, when is.auto.Dim = TRUE
.scaleAdjust	Logical value whether the PC projection is scaled to mean of zero and SD of 1
is.sparse	Logical value whether meta-eigenvector matrix is penalized to encourage sparseness
Lambda	Tuning parameter for penalty constant

Value

res = a list of meta analytic PCA
res$v = Meta eigenvector matrix
res$coord = Meta PC projection from input data

Author(s)

SungHwan Kim [e-mail]

Examples

library(yeastCC)
data(yeastCC)
data<-Biobase::exprs(yeastCC)
library(impute)
data.na<-is.na(data)
data.na.length<-apply(data.na, 1, sum)
data.sd<-apply(as.matrix(data), 1, sd, na.rm=TRUE)

new.data<-data[data.na.length<77*0.1 & data.sd>0.45,]    
Spellman <- list(alpha=impute.knn(new.data[,5:22])$data,
                cdc15=impute.knn(new.data[,23:46])$data,
                cdc28=impute.knn(new.data[,47:63])$data,
                elu=impute.knn(new.data[,64:77])$data)

## Standard Principal Component Analysis ##
Spellman.prcomp <- list(alpha=svd(t(Spellman$alpha))$v,
                       cdc15=svd(t(Spellman$cdc15))$v,
                       cdc28=svd(t(Spellman$cdc28))$v,
                       elu=svd(t(Spellman$elu))$v)

# res1 <- meta.pca(DList=Spellman, method="SV", Meta.Dim=2, is.auto.Dim = TRUE)
res2 <- meta.pca(DList=Spellman, method="SSC", Meta.Dim=2, is.auto.Dim = TRUE)

par(mfrow=c(6,4), cex=1 ,oma=c(0,6,2,0), mar=c(0.2,0.2,0.2,0.2))
for(i in 1:4) {
 for(j in 1:4) {
   .coord <- scale(t(Spellman[[j]]), scale=FALSE) 
   #.coord <- t(Spellman[[j]]-pcaPP::l1median(t(Spellman[[j]]),trace=-1)) %*% Spellman.prcomp[[i]]$loadings[,1:2]
   plot(.coord[,1], .coord[,2], type="n", xlab="", ylab="", xaxt="n", yaxt="n"
        ,ylim=c(min(.coord[,2])-1.5, max(.coord[,2])+1.5)
        ,xlim=c(min(.coord[,1])-1, max(.coord[,1])+1))
   text(.coord[,1], .coord[,2], 1:nrow(.coord), cex=1)
   lines(.coord[,1], .coord[,2], col='grey55')
 }
}

## SV
#for(j in 1:4) {
#  .coord <- res1$coord[[j]]
#  plot(.coord[,1], .coord[,2], type="n", xlab="", ylab="", xaxt="n", yaxt="n")
#  text(.coord[,1], .coord[,2], 1:nrow(.coord), cex=1)
#  lines(.coord[,1], .coord[,2])
#  }

## SSC
for(j in 1:4) {
 .coord <- res2$coord[[j]]
 plot(.coord[,1], .coord[,2], type="n", xlab="", ylab="", xaxt="n", yaxt="n"
      ,ylim=c(min(.coord[,2])-1.5, max(.coord[,2])+1.5)
      ,xlim=c(min(.coord[,1])-1, max(.coord[,1])+1))
 text(.coord[,1], .coord[,2], 1:nrow(.coord), cex=1)
 lines(.coord[,1], .coord[,2], col='grey55')
}


## Sparse MetaPCA (SSC)
library(PMA)
library(doMC)
res4 <- meta.pca(DList=Spellman, method="SSC", Meta.Dim=2, is.auto.Dim = TRUE, is.sparse=TRUE, Lambda=6)
res4 <- res4$v
coord <- list()
for(i in 1:4) {
 coord[[i]] <- t(Spellman[[i]]) 
}
for(j in 1:4) {
 .coord <- coord[[j]]
 plot(.coord[,1], .coord[,2], type="n", xlab="", ylab="", xaxt="n", yaxt="n"
      ,ylim=c(min(.coord[,2])-1.5, max(.coord[,2])+1.5)
      ,xlim=c(min(.coord[,1])-1, max(.coord[,1])+1))
 text(.coord[,1], .coord[,2], 1:nrow(.coord), cex=1)
 lines(.coord[,1], .coord[,2], col='grey55')     
}

####################################################################################################
## Searching the optimal tuning parameter based on the proportion of increased explained variance
####################################################################################################

optimal.lambda <- meta.pca.cv(DList=Spellman, method="SSC", Meta.Dim=2, CV_lambda = seq(1,10,1), is.plot=TRUE)
## optimal.lambda = 8

metaOmic/metaPCA documentation built on May 22, 2019, 6:54 p.m.