\documentclass[a4paper,12pt]{article}
%\VignetteIndexEntry{Manual for the phenoTest library}
%\VignettePackage{phenoTest,GSEABase}
\usepackage{amsmath} % need for subequations
\usepackage{amssymb} %useful mathematical symbols
\usepackage{bm} %needed for bold greek letters and math symbols
\usepackage{graphicx} % need for PS figures
\usepackage{hyperref} % use for hypertext links, including those to external documents and URLs
\usepackage{natbib} %number and author-year style referencing
\begin{document}
\title{Test association between phenotype and gene expression}
\author{Evarist Planet \\
\small{Bioinformatics \& Biostatistics Unit} \\
\small{IRB Barcelona}}
\date{} %comment to include current date
\maketitle
\tableofcontents
\section{Introduction}
\label{sec:intro}
Imagine a situation where we have gene expression and phenotype
variables and we want to test the association of each gene with
phenotype. We would probably be interested in testing association of
groups of genes (or gene sets) with phenotype. This library provides
the tools to do both things in a way that is efficient, structured,
fast and scalable. We also provide tools to do \emph{GSEA} (Gene set
enrichment analysis) of all phenotype variables at once.
The functions and methods presented on this \texttt{vignette} provide
tools to easily test association between gene expression levels of
individual genes or gene sets of genes and the selected phenotypes of
a given gene expression dataset. These can be particularly useful for
datasets arising from RNAseq or microarray gene expression studies.
We will load the \texttt{ExpressionSet} of a cohort (\emph{GSE2034})
we downloaded from \href{http://www.ncbi.nlm.nih.gov/geo/}{GEO}.
\begin{scriptsize}
<<loadLibAndData>>=
options(width=100)
library(phenoTest)
data(eset)
eset
@
\end{scriptsize}
For illustration purposes we selected the first 1000 genes and the
first 100 samples, created a continuous variable (named
\emph{Tumor.size}) and added a new category to the categorical
variable \emph{lymph.node.status} to illustrate the functionality of
the package.
\begin{scriptsize}
<<preprocess>>=
Tumor.size <- rnorm(ncol(eset),50,2)
pData(eset) <- cbind(pData(eset),Tumor.size)
pData(eset)[1:20,'lymph.node.status'] <- 'positive'
@
\end{scriptsize}
\section{Individual gene(s) association with phenotype(s)}
\label{sec:Preprocess}
\subsection{Creating an \texttt{epheno}}
The \texttt{epheno} object will contain the univariate association between a list of phenotype
variables and the gene expression from the given \texttt{ExpressionSet}.
We will use the \texttt{ExpressionPhenoTest} function to create the \texttt{epheno} object.
We will have to tell this function which phenotype variables we want to test and the type of
these variables (if they are \emph{ordinal}, \emph{continuous}, \emph{categorical} or \emph{survival}
variables).
For this purpose we will create a variable called (for instance) \texttt{vars2test}.
This variable has to be of class \texttt{list} with components \emph{continuous}, \emph{categorical},
\emph{ordinal} and \emph{survival} indicating which phenotype variables should be tested. \emph{continuous},
\emph{categorical} and \emph{ordinal} must be character vectors, \emph{survival} a matrix with columns
named \emph{time} and \emph{event}.
The names must match names in \texttt{names(pData(eset))} (being
\texttt{eset} the \texttt{ExpressionSet} of the cohort we are
interested in).
\begin{scriptsize}
<<setVars2test>>=
head(pData(eset))
survival <- matrix(c("Relapse","Months2Relapse"),ncol=2,byrow=TRUE)
colnames(survival) <- c('event','time')
vars2test <- list(survival=survival,categorical='lymph.node.status',continuous='Tumor.size')
vars2test
@
\end{scriptsize}
Now we have everything we need to create the \texttt{epheno} object:
\begin{scriptsize}
<<runExpressionPhenoTest>>=
epheno <- ExpressionPhenoTest(eset,vars2test,p.adjust.method='none')
epheno
@
\end{scriptsize}
P values can also be adjusted afterwards:
\begin{scriptsize}
<<pValueAdjust>>=
p.adjust.method(epheno)
epheno <- pAdjust(epheno,method='BH')
p.adjust.method(epheno)
@
\end{scriptsize}
The \texttt{epheno} object extends the \texttt{ExpressionSet} object and therefore methods that are
available for \texttt{ExpressionSet}s are also available for \texttt{epheno}s.
The effect of both \emph{continuous}, \emph{categorical} and \emph{ordinal} phenotype variables on gene
expression levels are tested via \texttt{lmFit} from package \texttt{limma} (\cite{smyth:2005}). For
\emph{ordinal} variables a single coefficient is used to test its effect on gene expression (trend
test), which is then used to obtain a P-value. Gene expression effects on \emph{survival} are tested
via Cox proportional hazards model (\cite{cox:1972}), as implemented in function \texttt{coxph} from
package \texttt{survival}.
If we want we can compute posterior probabilities instead of pvalues
we can set the argument \texttt{approach}='bayesian'.
The default value is 'frequentist'.
\texttt{ExpressionPhenoTest} implements parallel computing via the function \texttt{mclapply} from
the package \texttt{multicore}. Currently \texttt{multicore} only operates on Unix systems. If you
are a windows user you should set \texttt{mc.cores}=1 (the default).
\subsection{Useful methods for the \texttt{epheno} object}
Some of the methods for the epheno objects are shown here.
The object can be subseted by phenotype names:
\begin{scriptsize}
<<ephenoSubsetByName>>=
phenoNames(epheno)
epheno[,'Tumor.size']
epheno[,2]
@
\end{scriptsize}
or by class (class can be ordinal, continuous, categorical or survival):
\begin{scriptsize}
<<ephenoSubsetByClass>>=
phenoClass(epheno)
epheno[,phenoClass(epheno)=='survival']
@
\end{scriptsize}
\texttt{epheno} objects contain information summarizing the association between genes and
phenotypes. \texttt{getMeans} can be used to obtain the average expression for each group in
categorical and ordinal variables, as well as for categorized version of the continuous variables.
\begin{scriptsize}
<<ephenoGetMeans>>=
head(getMeans(epheno))
@
\end{scriptsize}
Here we see that tumor size has been categorized into 3 groups. The
number of categories can be changed with the argument
\texttt{continuousCategories} in the call to
\texttt{ExpressionPhenoTest}.
\texttt{epheno} objects also contain fold changes and hazard ratios (for survival variables). These
can be accessed with \texttt{getSummaryDif}, \texttt{getFc} and \texttt{getHr}.
\begin{scriptsize}
<<ephenoGetSummaries>>=
head(getSummaryDif(epheno))
head(getFc(epheno))
head(getHr(epheno))
@
\end{scriptsize}
\texttt{ExpressionPhenoTest} also computes P-values. \texttt{eBayes} from \texttt{limma} package is
used for continuous, categorical and ordinal phenotypes. A Cox proportional hazards likelihood-ratio
test is used for survival phenotypes. P-values can be accessed with \texttt{getSignif}. Notice that a
single P-value is reported for each phenotype variable. For categorical variables these corresponds
to the overall null hypothesis that there are no differences between groups.
\begin{scriptsize}
<<ephenoGetSignif>>=
head(getSignif(epheno))
@
\end{scriptsize}
We can also ask for the variables we sent to the \texttt{ExpressionPhenoTest} function:
\begin{scriptsize}
<<ephenoGetVars2test>>=
getVars2test(epheno)
@
\end{scriptsize}
\subsection{Export an \texttt{epheno}}
Functions \texttt{export2csv} and \texttt{epheno2html} can be used to export to a comma separated value
(csv) or an html file. The html file will have useful links to online databases that will provide information
about each known gene. For more information about how to use these functions and examples read their help manuals.
\section{Gene set(s) association with phenotype(s)}
Gene sets can be stored in a list object. Each element of the list will contain one gene set.
The names of the list will be the names of the gene sets. Here we select genes at random to build our
gene sets:
\begin{scriptsize}
<<getSignatures>>=
set.seed(777)
sign1 <- sample(featureNames(eset))[1:20]
sign2 <- sample(featureNames(eset))[1:50]
mySignature <- list(sign1,sign2)
names(mySignature) <- c('My first signature','Another signature')
mySignature
@
\end{scriptsize}
Gene sets can also be stored in gene set collection objects. From here
on all functions have methods for gene sets stored as \texttt{list}s,
\texttt{GeneSet}s or \texttt{GeneSetCollection}s. You can use the one
you feel more confortable with. We will work with
\texttt{GeneSetCollection}:
\begin{scriptsize}
<<makeGeneSets>>=
library(GSEABase)
myGeneSetA <- GeneSet(geneIds=sign1, setName='My first signature')
myGeneSetB <- GeneSet(geneIds=sign2, setName='Another signature')
mySignature <- GeneSetCollection(myGeneSetA,myGeneSetB)
mySignature
@
\end{scriptsize}
\subsection{Plots that use \texttt{epheno} as input}
\texttt{barplotSignifSignatures} will plot the percentage of
up regulated and down regulated genes that are statistically
significant in each signature. In our random selection of genes we did
not find any statistically significant genes. Therefore, and just to
show the plot we set the alpha value 0.99. The plot can be seen in
Figure \ref{fig:barplotSignifSignatures}.
\begin{scriptsize}
\setkeys{Gin}{width=0.6\textwidth}
<<label=barplotSignifSignatures,include=FALSE>>=
barplotSignifSignatures(epheno[,'lymph.node.status'],mySignature,alpha=0.99, ylim=c(0, 1))
@
\begin{figure}
\begin{center}
<<label=barplotSignifSignatures,fig=TRUE,echo=FALSE>>=
<<barplotSignifSignatures>>
@
\end{center}
\caption{\texttt{barplotSignifSignatures}: Number of diferentially
expressed genes in each gene set that are statistically
significant. P-values test for differences in each signature between
the number of up and down regulated genes.}
\label{fig:barplotSignifSignatures}
\end{figure}
\end{scriptsize}
By default \texttt{barplotSignifSignatures} performs a binomial test
(\texttt{binom.test} from package \texttt{stats}) for each signature
to test if the proportions of up regulated and down regulated genes
are different.
For example, Figure \ref{fig:barplotSignifSignatures} indicates that in the
first signature the proportion of up regulated genes is higher than
the proportion of down regulated genes. The second signature shows no
significant statistical differences.
Sometimes we want to compare the proportions of up and down regulated
genes in our signature with the proportions of up and down regulated
of all genes in the genome. In this case we may provide a reference
signature via the argument \texttt{referenceSignature}. When providing the
\texttt{referenceSignature} argument a chi-square test comparing the
proportion of up and down regulated genes in each signature with the
proportion in the reference set will be computed.
When a reference gene set is provided and parameter
\texttt{testUpDown} is \texttt{TRUE} (by default it is \texttt{FALSE})
the proportion of up regulated genes is compared with those of the
reference gene set. The same is done for down regulated genes.
\texttt{barplotSignatures} plots the average log2 fold change or
hazard ratio of each phenotype for each gene set. Figure
\ref{fig:barplotSignatures} shows an example of it.
\begin{scriptsize}
\setkeys{Gin}{width=0.6\textwidth}
<<label=barplotSignatures,include=FALSE>>=
barplotSignatures(epheno[,'Tumor.size'],mySignature, ylim=c(0,1))
@
\begin{figure}
\begin{center}
<<label=barplotSignatures,fig=TRUE,echo=FALSE>>=
<<barplotSignatures>>
@
\end{center}
\caption{\texttt{barplotSignatures}: Averge fold change or hazard ratio.}
\label{fig:barplotSignatures}
\end{figure}
\end{scriptsize}
We can also cluster our samples in two clusters based on the expression levels of one gene set of genes and
then test the effect of cluster on phenotypes. For \emph{ordinal} and
\emph{continuous} variables a Kruskal-Wallis Rank Sum test is used, for \emph{categorical} variables
a chi-square test is used and for \emph{survival} variables a Cox proportional hazards likelihood-ratio
test is used. The \texttt{heatmapPhenoTest} function can be used to this end. Its results can be seen
in Figure \ref{fig:heatmap} and \ref{fig:kaplan}.
\begin{scriptsize}
\setkeys{Gin}{width=0.6\textwidth}
<<label=heatmapPhenoTestHeat,include=FALSE>>=
pvals <- heatmapPhenoTest(eset,mySignature[[1]],vars2test=vars2test[1],heat.kaplan='heat')
@
\begin{scriptsize}
<<heatmapPhenoTestHeatPvals>>=
pvals
@
\end{scriptsize}
\begin{figure}
\begin{center}
<<label=heatmapPhenoTestHeat,fig=TRUE,echo=FALSE>>=
<<heatmapPhenoTestHeat>>
@
\end{center}
\caption{Heatmap produced with \texttt{heatmapPhenoTest} function. All
variables in \texttt{vars2test} that are of class \emph{logical}
will be plotted under the heatmap.}
\label{fig:heatmap}
\end{figure}
\end{scriptsize}
\begin{scriptsize}
\setkeys{Gin}{width=0.6\textwidth}
<<label=heatmapPhenoTestKaplan,include=FALSE>>=
pvals <- heatmapPhenoTest(eset,mySignature[[1]],vars2test=vars2test[1],heat.kaplan='kaplan')
@
\begin{scriptsize}
<<heatmapPhenoTestKaplanPvals>>=
pvals
@
\end{scriptsize}
\begin{figure}
\begin{center}
<<label=heatmapPhenoTest2,fig=TRUE,echo=FALSE>>=
<<heatmapPhenoTestKaplan>>
@
\end{center}
\caption{Kaplan-Meier produced with \texttt{heatmapPhenoTest} function.}
\label{fig:kaplan}
\end{figure}
\end{scriptsize}
\subsection{GSEA (Gene Set Enrichment Analysis)}
A popular way to test association between gene sets' gene expression and phenotype is GSEA (\cite{subramanian:2005}).
The main idea is to test the association between the gene set \emph{as a whole} and a phenotype.
Although GSEA and several extensions are already available in other \emph{Biconductor} packages,
here we implement a slightly different extension. Most GSEA-like approaches assess statistical
significance by permuting the values of the phenotype of interest. From a statisticall point of view
this tests the null hypothesis that no genes are associated with phenotype. However in many
applications one is actually interested in testing if the proportion of genes associated with
phenotype in the gene set is greater than that outside of the gene set. As a simple example, imagine
a cancer study where 25\% of the genes are differentially expressed. In this setup a randomly chosen
gene set will have around 25\% of differentially expressed genes, and classical GSEA-like approaches
will tend to flag the gene set as statistically significant. In contrast, our implementation will
tend to select only gene sets with more than 25\% of differentially expressed genes.
We will use the \texttt{gsea} method to compute \emph{enrichment
scores} (see \cite{subramanian:2005} for details about the enrichment
scores) and \emph{simulated enrichment scores} (by permuting the
selection of genes).
The \emph{simulated enrichment scores} are used
to compute P-values and FDR.
We can summarize the results obtained using the \texttt{summary} method.
The following chunk of code is an illustrative example of it:
\begin{scriptsize}
<<gsea>>=
my.gsea <- gsea(x=epheno,gsets=mySignature,B=1000,p.adjust='BH')
my.gsea
summary(my.gsea)
@
\end{scriptsize}
We receive one message for each phenotype we are testing.
We can produce plots as follows:
\begin{scriptsize}
<<gseaPlot>>=
plot(my.gsea)
@
\end{scriptsize}
This will produce two plots (one for \emph{enrichment score} and another for \emph{normalised enrichment
score}) for every phenotype and gene set (in our case 12 plots). Following code shows an example on
plotting only \emph{enrichment score} for variable \emph{Relapse} on the first gene set of genes. Plot
can be seen in Figure \ref{fig:plotGsea}.
\begin{scriptsize}
<<label=plotGseaPrepare,include=FALSE>>=
my.gsea <- gsea(x=epheno[,'Relapse'],gsets=mySignature[1],B=100,p.adjust='BH')
summary(my.gsea)
@
<<label=plotGseaEs,include=FALSE>>=
plot(my.gsea,es.nes='es',selGsets='My first signature')
@
\begin{figure}
\begin{center}
<<label=plotGsea,fig=TRUE,echo=FALSE>>=
<<plotGseaEs>>
@
\end{center}
\caption{GSEA plot.}
\label{fig:plotGsea}
\end{figure}
\end{scriptsize}
\texttt{gsea} can be used not only with \texttt{epheno} objects but also with objects of class \texttt{numeric}
or \texttt{matrix}. For more information read the \texttt{gsea} function help.
Following similar ideas to \cite{virtaneva:2001} we also implemented a
Wilcoxon test. This can be used instead of the permutation test which
can be slow if we use a lot of permutations and we can not use the
\texttt{multicore} package. The plot we will obtain will also be
different. Instead of plotting the \emph{enrichment scores} we will
plot the density function and the mean log2 fold change or hazard
ratio of the genes that belong to our gene set. This will allow us to
compare how similar/different from 0 the mean of our gene set is. The
plot using Wilcoxon test can be seen in Figure
\ref{fig:plotGseaWilcox}.
\begin{scriptsize}
<<plotGseaWilcoxPrepare>>=
my.gsea <- gsea(x=epheno[,'Relapse'],gsets=mySignature,B=100,test='wilcox',p.adjust='BH')
summary(my.gsea)
@
<<label=plotGseaWilcox,include=FALSE>>=
plot(my.gsea,selGsets='My first signature')
@
\begin{figure}
\begin{center}
<<label=plotGseaWilcox,fig=TRUE,echo=FALSE>>=
<<plotGseaWilcox>>
@
\end{center}
\caption{GSEA plot using Wilcoxon test.}
\label{fig:plotGseaWilcox}
\end{figure}
\end{scriptsize}
Notice that using a Wilcoxon test is conceptually very similar to the average gene set fold change
presented in figure \ref{fig:barplotSignatures}.
A current limitation of \texttt{gseaSignatures} is that it does not consider the existance of
dependence between genes in the gene set. This will be addressed in future versions. Nevertheless we
believe \texttt{gseaSignatures} is usefull in that it targets the correct null hypothesis that gene
set is as enriched as a randomly selected gene set, opposed to testing that there are no enriched
genes in the set as is done in GSEA.
\bibliographystyle{plainnat}
\bibliography{references}
\end{document}