Gene set testing
library(knitr) opts_chunk$set(fig.path=paste0("figure/", sub("(.*).Rmd","\\1",basename(knitr:::knit_concord$get('infile'))), "-"))
Gene set testing
Here, we will explore software for testing differential expression in a set of genes. These tests differ from the gene-by-gene tests we saw previously. Again, the gene set testing software we will use lives in the
We download an experiment from the GEO website, using the
getGEO function from the
library(GEOquery) g <- getGEO("GSE34313") e <- g[]
This dataset is hosted by GEO at the following link: http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE34313
The experiment is described in the paper by Masuno 2011.
Briefly, the investigators applied a glucocorticoid hormone to cultured human airway smooth muscle. The glucocorticoid hormone is used to treat asthma, as it reduces the inflammation response, however it has many other effects throughout the different tissues of the body.
The groups are defined in the
e$condition <- e$characteristics_ch1.2 levels(e$condition) <- c("dex24","dex4","control") table(e$condition)
By examining boxplots, we can guess that the data has already been normalized somehow, and on the GEO site the investigators report that they normalized using Agilent software.
We will subset to the control samples and the samples treated with dexamethasone (the hormone) after 4 hours.
boxplot(exprs(e), range=0) names(fData(e)) lvls <- c("control", "dex4") es <- e[,e$condition %in% lvls] es$condition <- factor(es$condition, levels=lvls)
The following lines run the linear model in
limma. We note that the top genes are common immune-response genes (CSF2, LIF, CCL2, IL6). Also present is FKBP5, a gene which regulates and is regulated by the protein which receives the glucocorticoid hormone.
library(limma) design <- model.matrix(~ es$condition) fit <- lmFit(es, design=design) fit <- eBayes(fit) tt <- topTable(fit, coef=2, genelist=fData(es)$GENE_SYMBOL) tt
We will use the ROAST method for gene set testing. We can test a single gene set by looking up the genes which contain a certain GO ID, and providing this to the
roast function. We will show how to get such lists of genes associated with a GO ID in the next chunk.
The roast function performs an advanced statistical technique, rotation of residuals, in order to generate a sense of the null distribution for the test statistic. The test statistics in this case is the summary of the scores from each gene. The tests are self-contained because only the summary for a single set is used, whereas other gene set tests might compare a set to all the other genes in the dataset, e.g., a competitive gene set test.
The result here tells us that the immune response genes are significantly down-regulated, and additionally, mixed up and down.
# Immune response idx <- grep("GO:0006955", fData(es)$GO_ID) length(idx) r1 <- roast(es, idx, design) # ?roast r1
Testing multiple gene sets
We can also use the
mroast function to perform multiple roast tests. First we need to create a list, which contains the indices of genes in the ExpressionSet for each of a number of gene sets. We will use the
org.Hs.eg.db package to gather the gene set information.
# biocLite("org.Hs.eg.db") library(org.Hs.eg.db) org.Hs.egGO2EG go2eg <- as.list(org.Hs.egGO2EG) head(go2eg)
The following code unlists the list, then gets matches for each Entrez gene ID to the index in the ExpressionSet. Finally, we rebuild the list.
govector <- unlist(go2eg) golengths <- sapply(go2eg, length) head(fData(es)$GENE) idxvector <- match(govector, fData(es)$GENE) table(is.na(idxvector)) idx <- split(idxvector, rep(names(go2eg), golengths)) go2eg[] fData(es)$GENE[idx[]]
We need to clean this list such that there are no
NA values. We also clean it to remove gene sets which have less than 10 genes.
idxclean <- lapply(idx, function(x) x[!is.na(x)]) idxlengths <- sapply(idxclean, length) idxsub <- idxclean[idxlengths > 10] length(idxsub)
The following line of code runs the multiple ROAST test. This can take about 3 minutes.
r2 <- mroast(es, idxsub, design) head(r2) r2 <- r2[order(r2$PValue.Mixed),]
We can use the
GO.db annotation package to extract the GO terms for the top results, by the mixed test.
# biocLite("GO.db") library(GO.db) columns(GO.db) keytypes(GO.db) GOTERM[[rownames(r2)]] r2tab <- select(GO.db, keys=rownames(r2)[1:10], columns=c("GOID","TERM","DEFINITION"), keytype="GOID") r2tab[,1:2]
We can also look for the top results using the standard p-value and in the up direction.
r2 <- r2[order(r2$PValue),] r2tab <- select(GO.db, keys=rownames(r2)[r2$Direction == "Up"][1:10], columns=c("GOID","TERM","DEFINITION"), keytype="GOID") r2tab[,1:2]
Again but for the down direction.
r2tab <- select(GO.db, keys=rownames(r2)[r2$Direction == "Down"][1:5], columns=c("GOID","TERM","DEFINITION"), keytype="GOID") r2tab[,1:2]
Methods within the limma package
Wu D, Lim E, Vaillant F, Asselin-Labat ML, Visvader JE, Smyth GK. "ROAST: rotation gene set tests for complex microarray experiments". Bioinformatics. 2010. http://www.ncbi.nlm.nih.gov/pubmed/20610611
Di Wu and Gordon K. Smyth, "Camera: a competitive gene set test accounting for inter-gene correlation" Nucleic Acids Research, 2012. http://nar.oxfordjournals.org/content/40/17/e133
Subramanian A1, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA, Paulovich A, Pomeroy SL, Golub TR, Lander ES, Mesirov JP, "Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles" Proc Natl Acad Sci U S A. 2005. http://www.ncbi.nlm.nih.gov/pubmed/16199517
Correlation within gene sets
William T. Barry, Andrew B. Nobel, and Fred A. Wright, "A statistical framework for testing functional categories in microarray data" Ann. Appl. Stat, 2008. http://projecteuclid.org/euclid.aoas/1206367822
William Barry has a package
safe in Bioconductor for gene set testing with resampling.
Daniel M Gatti, William T Barry, Andrew B Nobel, Ivan Rusyn and Fred A Wright, "Heading Down the Wrong Pathway: on the Influence of Correlation within Gene Sets", BMC Genomics, 2010. http://www.biomedcentral.com/1471-2164/11/574#B24
Gene sets and power
The following article points out an issue with gene set testing: the power to detect differential expression for an individual gene depends on the number of NGS reads which align to that gene, which depends on the transcript length among other factors.
Alicia Oshlack* and Matthew J Wakefield, "Transcript length bias in RNA-seq data confounds systems biology", Biology Direct, 2009. http://www.biologydirect.com/content/4/1/14
The dataset used in this lab
Masuno K, Haldar SM, Jeyaraj D, Mailloux CM, Huang X, Panettieri RA Jr, Jain MK, Gerber AN., "Expression profiling identifies Klf15 as a glucocorticoid target that regulates airway hyperresponsiveness". Am J Respir Cell Mol Biol. 2011. http://www.ncbi.nlm.nih.gov/pubmed/21257922