Introduction to Random Forests

for Gene Expression Data

Utah State University – Spring 2012

STAT 5570: Statistical Bioinformatics
Notes 3.4

1
References

 Breiman, Machine Learning (2001)

45(1): 5-32.
 Diaz-Uriarte and Alvarez de Andres,
BMC Bioinformatics (2006) 7:3.
 Cutler, Cutler, and Stevens (2009). In
Li and Xu, editors, High-Dimensional
Data Analysis in Cancer Research,
pages 83-101.
2
Gene Profiling / Selection
 “Observe” gene expression in different
conditions – healthy vs. diseased, e.g.

 Use simultaneous expression “profiles” of

thousands of genes (what are the genes
doing across arrays)

 Look at which genes are “important” in

“separating” the two conditions; i.e., what
determines the conditions’ “signatures”

3
Machine Learning
 Computational & statistical inference
processes:
observed data  reusable algorithms for prediction
 Why “machine”?
want minimal: human involvement
 Why “learning”?
develop ability to predict
 Here, supervised learning:
use knowledge of condition type

4
Machine Learning Methods
 Neural Network
 SVM (Support Vector Machine)
 RPART (Recursive PArtitioning and Regression
Trees)
 CART (Classification and Regression Trees)
 Ensembling Learning (average of many trees)
 Boosting (Shapire et al., 1998)
 Bagging (Breiman, 1996)
 RandomForests (Breiman, 2001; Cutler & Stevens 2006;
Cutler, Cutler, & Stevens 2008)

5
CART: Classification and Regression Trees
 Each individual (array) has data on many predictors
(genes) and one response (disease state)

 Think of a tree, with Diseased

(18:21)
splits based on levels
of specific predictors Gene1 > 8.3 Gene1 < 8.3

Healthy Diseased
 Choose predictors and (15:10) (3:11)
split levels to maximize
“purity” in new groups; the best split at each node

 Prediction made by: passing test cases down tree

6
CART generalized: Random Forests
 Rather than using all predictors and all individuals to
make a single tree, make a forest of many (ntree)
trees, each one based on a random selection of
predictors and individuals

 Each tree is fit using a bootstrap sample of data

(draw with replacement)
and ‘grown’ until each node is ‘pure’

 Each node is split using the best among a subset (of size
mtry) of predictors randomly chosen at that node
(default is sqrt. of # of predictors)
(special case using all predictors: bagging)

 Prediction made by aggregating across the forest

(majority vote or average)

7
How to measure “goodness”?
 Each tree fit on a “training” set (bootstrap
sample), or the “bag”

 The left-over cases (“out-of-bag”) can be

used as a “test” set for that tree
(usually 1/3 of original data)

 The “out-of-bag” (OOB) error rate is the:

% misclassification

8
What does RF give us?
 Kind of a “black box”
– but can look at “variable importance”
 For each tree, look at the OOB data:
 Permute values of predictor j among all OOB cases

 Pass OOB data down the tree, save the predictions

 For case i of OOB and predictor j , get:

OOB error rate with variable j permuted –
OOB error rate before permutation
 Average across forest to get overall variable
importance for each predictor j

9
Why “variable importance”?

 Recall: identifying conditions’ signatures

 Sort genes by some criterion

 Want smallest set of genes to achieve good

diagnostic ability

10
Recall – ALL Data
 RMA-preprocessed gene expression data
 Chiaretti et al., Blood (2004) 103(7)
 12625 genes (hgu95av2 Affymetrix GeneChip)
 128 samples (arrays)
 phenotypic data on all 128 patients, including:
 95 B-cell cancer

 33 T-cell cancer

11
ALL subset: 4,400 genes, 30 arrays
(15 B, 15 T)
Look at top 25 most important genes
from Random Forest.
Color scale:
purple (low) to orange (high)

12
memory.limit(size=4000)

# load data
library(affy); library(ALL); data(ALL)

# obtain relevant subset of data; similar to slide 11 of Notes 3.2

# (here, filter genes on raw scale, then return to log scale)
# also, keep 30 arrays here JUST for computational convenience
# (in-class)
library(genefilter); e.mat <- 2^(exprs(ALL)[,c(81:110)])
ffun <- filterfun(pOverA(0.20,100))
t.fil <- genefilter(e.mat,ffun)
small.eset <- log2(e.mat[t.fil,])
dim(small.eset) # 4400 genes, 30 arrays (15 B and 15 T)
group <- c(rep('B',15),rep('T',15)) # classification, in order

# One RF
library(randomForest)
set.seed(1234)
print(date())
rf <- randomForest(x=t(small.eset),y=as.factor(group),ntree=10000)
print(date()) # about 25 seconds

13
# Look at variable importance
imp.temp <- abs(rf$importance[,])
t <- order(imp.temp,decreasing=TRUE)
plot(c(1:nrow(small.eset)),imp.temp[t],log='x',cex.main=1.5,
xlab='gene rank',ylab='variable importance',cex.lab=1.5,
pch=16,main='ALL subset results')

# Get subset of expression values for 25 most 'important' genes

gn.imp <- names(imp.temp)[t]
gn.25 <- gn.imp[1:25] # vector of top 25 genes, in order
t <- is.element(rownames(small.eset),gn.25)
sig.eset <- small.eset[t,]
# matrix of expression values, not necessarily in order

## Make a heatmap, with group differences obvious on plot

library(RColorBrewer)
hmcol <- colorRampPalette(brewer.pal(11,"PuOr"))(256)
colnames(sig.eset) <- group # This will label the heatmap columns
csc <- rep(hmcol[50],30)
csc[group=='T'] <- hmcol[200]
# column side color will be purple for T and orange for B
heatmap(sig.eset,scale="row", col=hmcol,ColSideColors=csc)

14
 A Better Variable Importance Plot
varImpPlot(rf, n.var=25, main='ALL Subset Results')

imp.temp <- importance(rf)

t <- order(imp.temp,decreasing=TRUE)
gn.imp <- names(imp.temp)[t]

15
Can focus on “variable selection”
 Iteratively fit many forests, each time discarding predictors
with low importance from previous iterations

 Use bootstrap to assess standard error of error rates

 Choose the forest with the smallest number of genes

whose error rate is within u standard errors of the minimum
error rate of all forests (u = 0 or 1, typically)

 Reference: Diaz-Uriarte and Alvarez de Andres, BMC

Bioinformatics (2006) 7:3.

 Online tool: http://genesrf.bioinfo.cnio.es

16
RF Variable Selection on ALL subset

Subset: 30 arrays, 4400 genes

First forest: 10,000 trees; Subsequent forests: 2,000 trees
At each iteration, drop 20% of variables (genes) until a 2-variable model is
built; return the best of the series of models considered.

17
# Look at variable selection
library(varSelRF)
set.seed(1234)
print(date())
rfsel <- varSelRF(t(small.eset),as.factor(group),
ntree=10000, ntreeIterat=2000, vars.drop.frac=0.2)
# 50 seconds
print(date())
rf.sig.gn <- rfsel$selected.vars # "38147_at" "38319_at"

# Visualize these two genes

exp.gn.1 <- small.eset[rownames(small.eset)==rf.sig.gn[1],]
exp.gn.2 <- small.eset[rownames(small.eset)==rf.sig.gn[2],]
use.pch <- c(rep(1,15),rep(16,15)) # Define plotting characters
use.col <- c(rep(1,15),rep(2,15)) # Define plotting colors
plot(exp.gn.1,exp.gn.2,col=use.col,main='30 subset arrays',
cex.main=1.5, cex.lab=1.5, xlab=rf.sig.gn[1], ylab=rf.sig.gn[2],
pch=use.pch,cex=2)
legend('bottomright',
c('B-cell','T-cell'),pch=c(1,16),col=c(1,2),cex=1.5)

# Note that set.seed(123) leads to two other genes:

# "2059_s_at" "38319_at"
# Also: rfsel$firstForest is the same as the slide 13 rf object

18
# Did this overfit these 30 arrays?
# Look at the other 98 (first 80 are B-cell, last 18 are T-cell)
eset.2 <- exprs(ALL)[,c(1:80,111:128)]
group.2 <- c(rep(0,80),rep(1,18))
exp.gn.1 <- eset.2[rownames(eset.2)==rf.sig.gn[1],]
exp.gn.2 <- eset.2[rownames(eset.2)==rf.sig.gn[2],]
use.pch.2 <- c(rep(1,80),rep(16,18))
use.col.2 <- c(rep(1,80),rep(2,18))
plot(exp.gn.1,exp.gn.2,col=use.col.2, main='non-subset arrays',
cex.main=1.5,cex=2,cex.lab=1.5, xlab=rf.sig.gn[1],
ylab=rf.sig.gn[2], pch=use.pch.2)
legend('bottomright',
c('B-cell','T-cell'),pch=c(1,16),col=c(1,2),cex=1.5)

19
RF Variable Selection on ALL (full)

Each condition has a profile / signature

(across these genes)

full ALL data: 12,625 genes, 128 arrays 20

# RF variable selection with full data set

# set seed and define initial objects

set.seed(123)
eset <- exprs(ALL) # 12625 genes, 128 arrays
cell <- c(rep(0,95),rep(1,33))
# first 95 are B-cell; last 33 are T-cell

print(date())
rf.big <- varSelRF(t(eset),as.factor(cell),
ntree=10000, ntreeIterat=2000, vars.drop.frac=0.2)
print(date()) # about 14 minutes

rf.gn <- rf.big$selected.vars

# "33039_at" "35016_at" "38319_at"

21
# make scatterplot matrix, with points colored by cell type
rf.eset <- t(eset[is.element(rownames(eset),rf.gn),])
use.pch <- c(rep(1,95),rep(16,33))
use.col <- cell+1
pairs(rf.eset,col=use.col,pch=use.pch,cex=1.5)

# Now - make a profile plot

t <- is.element(rownames(eset),rf.gn)
rf.eset <- eset[t,]
library(MASS)
parcoord(t(rf.eset), col=cell+1, lty=cell+1, lwd=3, var.label=TRUE)
legend(1.2,.15,c('B-cell','T-cell'),lty=c(1,2),
lwd=3,col=c(1,2),bty='n')

22
Summary: RF for gene expression data
 Works well even with: many more variables than
observations, many-valued categoricals, extensive
missing values, badly unbalanced data
 Low error (comparable w/ boosting and SVM)
 Robustness (even with large “noise” in genes)
 Does not overfit
 Fast, and invariant to monotone transformations of
the predictors
 Free! (Fortran code by Breiman and Cutler)
 Returns the importance of predictors (gene)
 Little need to tune parameters
 But, no: statistical inference

23