In parsifal9/RFlocalfdr: Significance Level for Random Forest Impurity Importance Scores

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
  )
knitr::opts_chunk$set(fig.width=6, fig.height=6, dpi=300,echo = FALSE)
knitr::opts_chunk$set(fig.pos = "H", out.extra = "")

A small example using the smoking data set, containing normalized transcript measurements for 51 subjects (23 "never-smoked" and 34 "smokers") and 22283 transcripts of lung tissue. See @Spira.et.al.2004

I have removed that data set temporarily until I solve a problem with upload the package to CRAN

load the data

library(RFlocalfdr.data)
data(smoking)
?smoking 
y<-smoking$y
smoking_data<-smoking$rma
y.numeric <-ifelse((y=="never-smoked"),0,1)

if (!require("BiocManager", quietly = TRUE))
    install.packages("BiocManager")

BiocManager::install("GEOquery")
BiocManager::install("annotate")

install.packages("geneExpressionFromGEO")
library(geneExpressionFromGEO)
DF1 <- getGeneExpressionFromGEO("GSE994", FALSE, FALSE)  #retrieveGeneSymbols, verbose = FALSE)

aa<-match(gsub("([A-Z0-9]*).*","\\1", rownames(smoking_data)), colnames(DF1))
all.equal(aa,1:57)
#[1] TRUE

temp <- t(DF1[1:57])
#but class temp is a data.frame not a ‘AffyBatch’ object. So how do we normalize it?

## library(affy)
## rma.data <- rma(data)
## or
## rma.data <- expresso(data,
## bgcorrect.method = "rma",
## normalize.method = "quantiles",
## pmcorrect.method = "pmonly",
## summary.method = "medianpolish")

fit a ranger model

library(ranger)
rf1 <-ranger(y=y.numeric ,x=smoking_data,importance="impurity",seed=123, num.trees = 10000,
             classification=TRUE)
t2 <-count_variables(rf1)
imp<-log(rf1$variable.importance)
#png("./supp_figures/smoking_log_importances.png")
plot(density(imp),xlab="log importances",main="")
#dev.off()

# starting from a randomForest model
library(RFlocalfdr)
library(randomForest)
sert.seed(123)
rf2 <-randomForest(y=factor(y.numeric) ,x=smoking_data,importance=TRUE, ntree = 10000)
imp.rf2 <- log(rf2$importance[,"MeanDecreaseGini"])
t2.rf2 <-varUsed(rf2)
plot(density(imp.rf2),xlab="log importances",main="")

cutoffs <- c(2,3,4,5)
res.con<- determine_cutoff(imp.rf2,t2.rf2 ,cutoff=cutoffs,plot=c(2,3,4,5))

plot(cutoffs,res.con[,3],pch=15,col="red",cex=1.5,ylab="max(abs(y - t1))")
cutoffs[which.min(res.con[,3])]
#2

```r, and each block consists of correlated (identical variables).", fig.align="center", out.width = '50%'} knitr::include_graphics("./supp_figures/smoking_log_importances.png")

# Determine a cutoff to get a unimodal density.

See \@ref(fig:log_importances) for the log importances. They are clearly multimodal and we try to determine a cutoff so
that we are left with a unimodal distribution.

```r
cutoffs <- c(2,3,4,5)
#png("./supp_figures/smoking_data_determine_cutoff.png")
res.con<- determine_cutoff(imp,t2,cutoff=cutoffs,plot=c(2,3,4,5))
#dev.off()

#png("./supp_figures/smoking_data_determine_cutoffs_2.png")
plot(cutoffs,res.con[,3],pch=15,col="red",cex=1.5,ylab="max(abs(y - t1))")
#dev.off()
cutoffs[which.min(res.con[,3])]

fit RFlocalfdr

We select a cutoff of 3 and fit the RFlocalfdr model

temp<-imp[t2 > 3]
temp <- temp - min(temp) + .Machine$double.eps
qq <- plotQ(temp,debug.flag = 1)
ppp<-run.it.importances(qq,temp,debug.flag = 0)


#png("./supp_figures/smoking_significant_genes.png")
#aa<-significant.genes(ppp,temp,cutoff=0.05,do.plot=1)
aa<-significant.genes(ppp,temp,cutoff=0.05,debug.flag=0,do.plot=TRUE,use_95_q=TRUE)
#dev.off()
length(aa$probabilities) # 17

aa<-significant.genes(ppp,temp,cutoff=0.05,debug.flag=0,do.plot=TRUE,use_95_q=FALSE)
length(aa$probabilities) # 19

The option do.plot=1 returns a plot containing the

• histogram of the importances in magenta
• fdr in black -- axis scale on right
• alpha value for significance -- axis scale on right

The option do.plot=2 returns the same plot with the addition of

• fitted curve using estimates_C_0.95 in red
• 0.95 quantile of the fitted distribution, also in red
• cutoff for significant genes (in orange)
• abline(v = object$C_0.95, lwd = 2, col = "blue") what are these
• abline(v = object$cc, lwd = 2, col = "purple")

sessionInfo()

devtools::install_github("parsifal9/RFlocalfdr", build_vignettes = TRUE, force = TRUE)

library(RFlocalfdr)
data(smoking)
?smoking 
y<-smoking$y
smoking_data<-smoking$rma
y.numeric <-ifelse((y=="never-smoked"),0,1)

library(ranger)
rf1 <-ranger(y=y.numeric ,x=smoking_data,importance="impurity",seed=123, num.trees = 10000,
             classification=TRUE)
t2 <-count_variables(rf1)
imp<-log(rf1$variable.importance)
#png("./supp_figures/smoking_log_importances.png")
plot(density(imp),xlab="log importances",main="")
#dev.off()



cutoffs <- c(2,3,4,5)
#png("./supp_figures/smoking_data_determine_cutoff.png")
res.con<- determine_cutoff(imp,t2,cutoff=cutoffs,plot=c(2,3,4,5))
#dev.off()

#png("./supp_figures/smoking_data_determine_cutoffs_2.png")
plot(cutoffs,res.con[,3],pch=15,col="red",cex=1.5,ylab="max(abs(y - t1))")
#dev.off()
cutoffs[which.min(res.con[,3])]


temp<-imp[t2 > 3]
temp <- temp - min(temp) + .Machine$double.eps
qq <- plotQ(temp,debug.flag = 1)
ppp<-run.it.importances(qq,temp,debug.flag = 0)


#png("./supp_figures/smoking_significant_genes.png")
#aa<-significant.genes(ppp,temp,cutoff=0.05,do.plot=1)
aa<-significant.genes(ppp,temp,cutoff=0.05,debug.flag=0,do.plot=TRUE,use_95_q=TRUE)
#dev.off()
length(aa$probabilities) # 17  -- Roc gets 30

aa<-significant.genes(ppp,temp,cutoff=0.05,debug.flag=0,do.plot=TRUE,use_95_q=FALSE)
length(aa$probabilities) # 19-- Roc gets 30

parsifal9/RFlocalfdr documentation built on Feb. 3, 2025, 7:58 p.m.