Estrogen Data
A 2x2 factorial experiment

Gordon Smyth
16 August 2005

1. Aims

This case study illustrates more advanced linear modeling with Affymetrix single-channel microarrays. The popular 2x2 factorial design is considered. Use of Bioconductor annotation for Affymetrix arrays is illustrated. The case study goes on to consider significance tests using gene sets.

2. Required data

The estrogen data set is required for this lab and can be obtained from estrogen.zip. You should create a clean directory, unpack this file into that directory, then set that directory as your working directory for your R session using setwd() or otherwise. By typing dir() you should see eight .CEL files and three text files. On my computer, I see:

> setwd("C:/Gordon/www/bioinf/marray/bioc2005/_src/estrogen")
> dir()
 [1] "estrogen.txt"         "high10-1.cel"         "high10-2.cel"        
 [4] "high48-1.cel"         "high48-2.cel"         "knownERgenes.txt"    
 [7] "low10-1.cel"          "low10-2.cel"          "low48-1.cel"         
[10] "low48-2.cel"          "predictedERgenes.txt"

To repeat this case study in full you will need to have the R packages affy, hgu95av2cdf, hgu95av2 and xtable installed.

3. Estrogen experiment

The data gives results from a 2x2 factorial experiment on MCF7 breast cancer cells using Affymetrix HGU95av2 arrays. The factors in this experiment were estrogen (present or absent) and length of exposure (10 or 48 hours). The aim of the study is the identify genes which respond to estrogen and to classify these into early and late responders. Genes which respond early are putative direct-target genes while those which respond late are probably downstream targets in the molecular pathway.

This data is from the estrogen data package on the Bioconductor website http://www.bioconductor.org/data/experimental.html. Rather than loading the data package here we simply using the nine basic data files from that package, in order to save storage space and to more closely mimic a real data analysis situation. The data set is discussed further by Scholtens [1,2] and in the Limma User's Guide.

4. Read the data

The first step in most analyses is to read the targets file which describes what RNA target has been hybridized to each array and, equally importantly, gives the names of the corresponding data files. The targets file is usually tab-delimited, but here it is white-space delimited.

 library(limma)
 targets <- readTargets("estrogen.txt", sep="")
 targets

You should see

> targets
      filename estrogen time.h
1  low10-1.cel   absent     10
2  low10-2.cel   absent     10
3 high10-1.cel  present     10
4 high10-2.cel  present     10
5  low48-1.cel   absent     48
6  low48-2.cel   absent     48
7 high48-1.cel  present     48
8 high48-2.cel  present     48

Now read the CEL file data into an AffyBatch object and normalize using RMA:

 library(affy)
 library(hgu95av2cdf)
 abatch <- ReadAffy(filenames=targets$filename)
 eset <- rma(abatch)

Here eset is a data object of class exprSet.

It is usual and appropriate to check data quality before continuing your analysis. Due to brevity we will skip over this in this lab. A full set of quality assessment plots can be found at http://www.stat.berkeley.edu/~bolstad/PLMImageGallery/ under the title "estrogen". These plots show no significant quality problems with any arrays in this dataset.

5. Create a design matrix

We have four pairs of replicate arrays so we should estimate four parameters in the linear model. There are many valid ways to choose a design matrix, but perhaps the simplest is to make each column correspond to a particular treatment combination:.

estrogenDesignMatrix.png

The four columns of the matrix correspond to absent10, present10, absent48 and present48, respectively. Another way to specify the design matrix is described in the Limma User's Guide.

This design matrix given above can be computed in R as follows:

 f <- paste(targets$estrogen,targets$time.h,sep="")
 f <- factor(f)
 f
 design <- model.matrix(~0+f)
 colnames(design) <- levels(f)
 design

6. Fit the linear model

Now that we have defined our design matrix, fitting a linear model is as simple as:

 fit <- lmFit(eset, design)

fit is an object of class MArrayLM.

 names(fit)

The fitted coefficents fit$coef from the model fit are just the mean log-expression for each treatment combination for each probe set. For this reason, this choice of design matrix is called the group means parametrization in the Limma User's Guide.

7. Define a contrast matrix

The idea now is to use contrasts to make any comparisons of interest between the four treatment combination. Contrasts are linear combinations of parameters from the linear model fit.

Contrasts.png

where beta.png is a vector of contrasts for gene g.png, C.png is the contrasts matrix, and alpha.png is a vector of coefficients (estimated log fold changes), obtained from a linear model fit.

We will estimate three contrasts (so our contrasts matrix will have three columns). The first contrast is an estrogen effect at time 10 hours, the second as an estrogen effect at time 48 hours and the third is the time effect in the absence of estrogen. These are not all the comparisons which might have been made.

estrogenContrastsMatrix.png
 cont.matrix <- makeContrasts(E10="present10-absent10",E48="present48-absent48",Time="absent48-absent10",levels=design)
 cont.matrix

8. Extract the linear model fit for the contrasts

 fit2  <- contrasts.fit(fit, cont.matrix)
 fit2  <- eBayes(fit2)

9. Assessing differential expression

We now use the function topTable to obtain a list genes differentially expressed between Estrogen-Present and Estrogen-Absent at time 10 hours, followed by a list of genes differentially expressed between Estrogen-Present and Estrogen-Absent at time 48 hours.

 colnames(fit2)
 topTable(fit2,coef=1)
 topTable(fit2,coef=2,adjust="fdr")
 topTable(fit2,coef=2)
 topTable(fit2,coef=2,adjust="fdr")

The function decideTests() provides a variety of ways to assign statistical significance to the contrasts while controlling for multiple testing.

 results <- decideTests(fit2)
 summary(results)
 vennDiagram(results)

10. Linking the gene lists to annotation information on the Internet

This section was prepared by James Wettenhall.

If the genes int the topTable have standard IDs (e.g. UniGene or GenBank), then they can be linked with external annotation information on the Internet. Load the annotation package hgu95av2, which can be obtained from http://www.bioconductor.org/data/metaData.html.

 library(hgu95av2cdf)
 library(hgu95av2)

Now we obtain:

 geneIDs <- ls(hgu95av2cdf)

It is possible to extract the annotation information from the appropriate R environments within the annotation package (hgu95av2), and store it in simple R data structures as follows:

 # geneSymbols <- unlist(as.list(hgu95av2SYMBOL))
 # geneNames <- unlist(as.list(hgu95av2GENENAME))
 # unigene <- unlist(as.list(hgu95av2UNIGENE))

However we must be very careful, because in recent versions of Bioconductor annotation packages, some gene IDs can map to multiple gene names, multiple symbols and/or multiple unigene clusters. The method above, while appearing simple, also does not allow for the possibility that the gene names are stored in a different order in the annotation package from the gene IDs. We will therefore use some rather complicated-looking R code to extract the gene symbols, names and unigene IDs from the hgu95av2 environment. Multiple entries will be collapsed and separated by semicolons, e.g. if geneID 001 corresponds to gene names "Name1" and "Name2", these will be collapsed into "Name1; Name2".

 geneSymbols <- as.character(unlist(lapply(mget(geneIDs,env=hgu95av2SYMBOL),
     function (symbol) { return(paste(symbol,collapse="; ")) } )))
 geneNames <- as.character(unlist(lapply(mget(geneIDs,env=hgu95av2GENENAME),
     function (name) { return(paste(name,collapse="; ")) } )))
 unigene <- as.character(unlist(lapply(mget(geneIDs,env=hgu95av2UNIGENE),
     function (unigeneID) { return(paste(unigeneID,collapse="; ")) } )))

The key functions to note in the code above are mget which extracts multiple annotation strings from the appropriate annotation environmentand lapply which applies our multiple-entry-collapsing function to every element in a list of annotation strings.

Now we abbreviate the gene names to a maximum length of 40 characters, for neat formatting in a table, and we extract the unigene ID from the string which contains "Hs" (for Homo sapiens) as well as the unigene ID.

 geneNames <- substring(geneNames,1,40)
 unigene <- gsub("Hs\\.","",unigene)

 genelist <- data.frame(GeneID=geneIDs,GeneSymbol=geneSymbols,GeneName=geneNames,
   UniGeneHsID=paste("<a href=http://www.ncbi.nlm.nih.gov/UniGene/clust.cgi?ORG=Hs&CID=",
   unigene,">",unigene,"</a>",sep=""))

Now we recreate the toptable for the two contrasts considered earlier, E10="present10-absent10" and E48="present48-absent48" this time providing a hyperlink to the UniGene website for each gene in the toptable.

 unigeneTopTableEst10 <- topTable(fit2,coef=1,n=20,genelist=genelist)
 unigeneTopTableEst48 <- topTable(fit2,coef=2,n=20,genelist=genelist)
 library(xtable)
 xtableUnigeneEst10 <- xtable(unigeneTopTableEst10,display=c("d","s","s","s","s","g","g","g","e","g"))
 xtableUnigeneEst48 <- xtable(unigeneTopTableEst48,display=c("d","s","s","s","s","g","g","g","e","g"))

 cat(file="estrogenUniGeneE10.html","<html>\n<body>")
 print.xtable(xtableUnigeneEst10,type="html",file="estrogenUniGeneE10.html",append=TRUE)
 cat(file="estrogenUniGeneE10.html","</body>\n</html>",append=TRUE)

 cat(file="estrogenUniGeneE48.html","<html>\n<body>")
 print.xtable(xtableUnigeneEst48,type="html",file="estrogenUniGeneE48.html",append=TRUE)
 cat(file="estrogenUniGeneE48.html","</body>\n</html>",append=TRUE)

The display argument to the xtable function is used to specify the format of text or numbers displayed in cells of the HTML table:

Format code Meaning
d Decimal (base ten) integer, e.g. 48
s Character string, e.g. "Block"
g General real floating-point number, e.g. 8.25
e Floating point number in exponent format, e.g. 1.02E-05

11. Gene Set Enrichment

This section was prepared by James Wettenhall.

11.1 Introduction to Gene Set Analysis

In this lab, we move beyond the analysis of individual genes, and consider sets of genes in microarray experiments. Another approach is to form gene sets based on a priori knowledge of common biological features shared by the genes. We consider a particular approach called gene set enrichment. We begin with a known set of genes and then test whether this set as a whole is differentially expressed in a microarray experiment. This type of test is useful when comparing one's microarray data with that of previous authors who have performed similar microarray experiments, because the lists of most differentially expressed genes reported by the previous authors can be regarded as a "gene set" and tested to determine whether the genes are also differentially expressed in the current context.

Gene set testing was introduced by Mootha et al [5] and Lamb et al [6] in 2003. Mootha et al define the concept of a gene set enrichment test. For a given set of genes, one can test whether the set as a whole is up-regulated, down-regulated or differentially expressed with individual genes possibly going in either direction. Sometimes performing the traditional differential expression analysis of individual genes will yield no statistically significant results, but there may be stronger evidence for differential expression of gene sets.

Now we turn our attention to tests for differential expression involving a set of genes. Mootha et al. [5] and Lamb et al. [6] made this methods popular in 2003. We will use a "gene set enrichment test", which is closely based on the one defined by Mootha et al. The gene set test can be used to test whether previous author's lists of differentially expressed genes are also differentially expressed in a current experiment similar to that of the previous authors. Another possible application is to try to find differential expression in microarray experiments which show no strong differential expression when testing for individual differentially expressed genes, but they might show more evidence of differential expression when testing a predefined set of genes. Defining a useful gene set for this sort of analysis is not always trivial. One possibility is to use a set of genes which share common gene ontologies, i.e. choose a set of genes which are all associated with GOs below a certain node in the GO DAG (Directed Acyclic Graph). We will begin with some artificial examples to illustrate the concept of gene set tests with a small number of made-up t-statistics. Then we will use two sets of genes thought to be regulated by the Estrogen Receptor (ERalpha) to demonstrate testing for differential expression of gene sets in the Estrogen data set.

The geneSetTest function in the limma package [8] is described in its help file, reproduced below:

This is essentially a stream-lined approach to Gene Set Enrichment Analysis introduced by Mootha et al (2003). Usually, 'statistics' is intended to hold t-like statistics, meaning that the genewise null hypotheses would be rejected for large positive or large negative values. Then 'alternative="greater"' can be used to test whether genes in the set tend to be up-regulated, 'alternative="less"' can be used to test whether the gene set is down-regulated, while 'alternative="two.sided"' tests whether the gene set holds highly ranked genes without regard to direction of change. Important note: if 'statistics' is an F-like statistic for which only large values are relevant for rejecting the null hypothesis, then you must use 'alternative="greater"' to get meaningful results.

We now demonstrate the use of the geneSetTest function in limma using a miniscule set of artifical made-up t-statistics, where as usual, a positive t statistic means that a gene is up-regulated (i.e. expressed more highly in a condition of interest), whilst a negative value means that the gene is down-regulated. A t-statistic close to zero means that the gene is not differentially expressed.

In the first example we use, the artificial t-statistics will range from -9 to 9, and we will select a set of three genes for the test, those with t-statistics of -8, -6 and -5, i.e. we will use the 2nd, 4th and 5th t-statistics from our vector of artificial t-statistics. If these t-statistics represented real genes, all three genes would show strong evidence of differential expression (down-regulation) individually, so they should certainly show strong evidence of differential expression as a set as well. The value returned by limma's geneSetTest function is a p-value.

 library(limma)
 sel <- c(2,4,5)
 stat <- -9:9
 stat[sel]
 geneSetTest(sel,stat,nsim=100)
 geneSetTest(sel,stat,ranks.only=TRUE)

If we did a test for up-regulation of the set, we would expect a large p-value (low evidence of up-regulation):

 geneSetTest(sel,stat,alternative="greater",nsim=100)

On the otherhand, a test for down-regulation should give a small p-value:

 geneSetTest(sel,stat,alternative="less",nsim=100)

11.2 Gene Set Analysis Example

We will again use the Estrogen data set through the fit2 linear model fit object.

We will use two sets of genes which are thought to be ER-regulated, i.e. regulated by the Estrogen Receptor alpha. The first set (Jin et al [4]) contains genes which have been experimentally verified to be ER-regulated and the second set (O'lone et al [7]) contains a large list of genes which are predicted to be ER-regulated by a model (so they may or may not be ER-regulated).

These gene sets (particularly the first one) should be differentially expressed between the breast cancer cells with estrogen reintroduced and the serum-starved breast cancer cells with no estrogen, because in the cells reintroduced to estrogen, the estrogen receptors (ERs) will bind the estrogen and as a result become activated, gaining the ability to regulate gene expression in the cells, hence resulting in differential expression between the cells with and without estrogen.

The data required for this exercise is available from knownERgenes.txt and predictedERgenes.txt. Read the gene lists into R:

 known <- read.delim("knownERgenes.txt",as.is=TRUE)
 knownERgenes <- known$UGCluster
 predicted <- read.delim("predictedERgenes.txt",as.is=TRUE)
 predictedERgenes <- predicted$UGCluster
 library(hgu95av2)
 unigene <- unlist(as.list(hgu95av2UNIGENE))
 knownERgenesOnChip <- match(knownERgenes,unigene)
 knownERgenesOnChip <- knownERgenesOnChip[!is.na(knownERgenesOnChip)]
 predictedERgenesOnChip <- match(predictedERgenes,unigene)
 predictedERgenesOnChip <- predictedERgenesOnChip[!is.na(predictedERgenesOnChip)]

Now we will use the moderated t-statistics calculated previously for the comparison of estrogen present and estrogen absent 10 hours after the estrogen was reintroduced into the cells. We try all three types of tests - "two-sided", "greater" (for genes up-regulated in the sample with estrogen present), and "less" (for genes down-regulated in the sample with estrogen present). We expect to get better (smaller) p-values for the known ER-regulated genes than for the predicted ER-regulated genes. The recent review of Estrogen-repressed genes in breast cancer cells by Zubairy and Oesterreich [9] suggests that the majority of genes regulated by estrogen receptors are actually repressed (down-regulated), so we should expect a lower p-value for the "less" test (at least for the known ER-regulated genes).

geneSetTest(knownERgenesOnChip,fit2$t[,1],"two.sided")
geneSetTest(knownERgenesOnChip,fit2$t[,1],"greater")
geneSetTest(knownERgenesOnChip,fit2$t[,1],"less")

set.seed(0)
geneSet <- sample(predictedERgenesOnChip,length(knownERgenesOnChip))
geneSetTest(geneSet,fit2$t[,1],"two.sided")
geneSetTest(geneSet,fit2$t[,1],"greater")
geneSetTest(geneSet,fit2$t[,1],"less")

geneSet <- sample(predictedERgenesOnChip,length(knownERgenesOnChip))
geneSetTest(geneSet,fit2$t[,1],"two.sided")
geneSetTest(geneSet,fit2$t[,1],"greater")
geneSetTest(geneSet,fit2$t[,1],"less")

geneSet <- sample(predictedERgenesOnChip,length(knownERgenesOnChip))
geneSetTest(geneSet,fit2$t[,1],"two.sided")
geneSetTest(geneSet,fit2$t[,1],"greater")
geneSetTest(geneSet,fit2$t[,1],"less")

Acknowledgements

Thanks to Denise Scholtens and Robert Gentleman for providing the estrogen data http://www.bioconductor.org/data/experimental.html. Thanks to Ben Bolstad for a quality analysis of this data.

Thanks to Hui Tang and Terry Speed for finding the known and predicted ER-regulated gene sets (Jin et al. [4] and O'lone et al. [7]). Thanks to James Wettenhall for material on linking to internet annotation and for the gene set worked example.

References on factorial studies

  1. Scholtens D, Gentleman R. Estrogen 2x2 Factorial Design. http://www.bioconductor.org/repository/devel/vignette/factDesign.pdf
  2. Scholtens D, Miron A, Merchant FM, Miller A, Miron PL, Iglehart JD, Gentleman R. Analyzing Factorial Designed Microarray Experiments. Journal of Multivariate Analysis. To appear.
  3. Smyth, G. K. (2005). Limma: linear models for microarray data. In: Bioinformatics and Computational Biology Solutions using R and Bioconductor, R. Gentleman, V. Carey, S. Dudoit, R. Irizarry, W. Huber (eds.), Springer, New York.
  4. Smyth, G. K., Thorne, N. P. and Wettenhall J. (2005) limma: Linear Models for Microarray Data User's Guide. http://bioinf.wehi.edu.au/limma (See the estrogen case study.)

References on gene sets

  1. Jin VX, Leu YW, Liyanarachchi S, Sun H, Fan1 M, Nephew1 KP, Huang T.H. and Davuluri DV (2004). Identifying estrogen receptor {alpha} target genes using integrated computational genomics and chromatin immunoprecipitation microarray. Nucleic Acids Research 32(22): 6627-6635
  2. Lamb J, Ramaswamy S, Ford HL, Contreras B, Martinez RV, Kittrell FS, Zahnow CA, Patterson N, Golub TR, Ewen ME (2003). A mechanism of cyclin D1 action encoded in the patterns of gene expression in human cancer. Cell; 114(3):323-34.
  3. Mootha VK, Lindgren CM, Eriksson KF, Subramanian A, Sihag S, Lehar J, Puigserver P, Carlsson E, Ridderstråle, M, Laurila E, Houstis N, Daly MJ, Patterson N, Mesirov JP, Golub TR, Tamayo P, Spiegelman BM, Lander ES, Hirschhorn JN, Altshuler D, and Groop LC (2003). PGC-1alpha Responsive Genes Involved in Oxidative Phosphorylation are Coordinately Downregulated in Human Diabetes, Nature Genetics 34(3):267-73.
  4. O'lone R, Frith MC, Karlsson EK, Hansen U (2004). Genomic Targets of Nuclear Estrogen Receptors. Molecular Endocrinology 18(8): 1859-1875
  5. Zubairy S, Oesterreich S (2005). Estrogen-repressed genes -- key mediators of estrogen action? Breast Cancer Res. 2005;7(4):163-4.

Return to list of Exercises

Fred Hutchinson Cancer Research Center