% \VignetteIndexEntry{ConsensusClusterPlus Tutorial}
% \VignettePackage{ConsensusClusterPlus}
\documentclass{article}
\usepackage{graphicx}
\begin{document}
\title{ConsensusClusterPlus (Tutorial)}
\author{ Matthew D. Wilkerson}
\maketitle
\SweaveOpts{keep.source=TRUE}

\section{Summary}
\texttt{ConsensusClusterPlus} is a tool for unsupervised class discovery.  This document provides a tutorial of how to use \texttt{ConsensusClusterPlus}.

\section{Brief description of \textit{Consensus Clustering} }
\textit{Consensus Clustering} \cite{monti} is a method that provides quantitative evidence for determining the number
 and membership of possible clusters within a dataset, such as microarray gene expression.  This method has gained popularity
 in cancer genomics, where new molecular subclasses of disease have been discovered \cite{hayes,verhaak}.  
 The \textit{Consensus Clustering} method involves subsampling from a set of items, such as microarrays, and determines clusterings of specified cluster counts (\textit{k}).  Then, pairwise \textit{consensus} values, the proportion that two items occupied the same cluster out of the number of 
times they occurred in the same subsample, are calculated and stored in a symmetrical \textit{consensus matrix} for each \textit{k}.  
The \textit{consensus matrix} is summarized in several graphical summaries that enable a user to decide upon a reasonable cluster number and membership.  
A web-based version of \textit{Consensus Clustering} is publicly available \cite{gp}.  For a formal description, see \cite{monti}.
\\
\\
\texttt{ConsensusClusterPlus} implements the \textit{Consensus Clustering} method in \textit{R} and extends it with
new features and graphical outputs that can aid users in class discovery.

\section{Tutorial}
There are three main steps to use \texttt{ConsensusClusterPlus}: preparing input data, running the program, and generating cluster-consensus and item-consensus.


\subsection{Preparing input data}
The first step is to gather some data for cluster analysis. These data could be the result of an experiment such as a mRNA expression microarray or
immunohistochemical staining intensities.  The input data format is a matrix where columns are samples (items), rows are features and cells are numerical values.
For this tutorial, we use the ALL gene expression data from the ALL library.  You can see the matrix d is already in the proper format.  The column and row names, which 
correspond to the sample and gene names, will be maintained in the output.
<<>>=
library(ALL)
data(ALL)
d=exprs(ALL)
d[1:5,1:5]
@

For the purpose of selecting the most informative genes for class detection, 
we reduce the dataset to the top 5,000 most variable genes, measured by median absolute deviation.
The choice of 5,000 genes and MAD can be substituted with other statistical variability filters.  
Users can decide what type of filtering to use or to skip filtering.  Another choice would be to supply weights for sampling genes see weightsFeatures in Additional Options.
<<>>=
mads=apply(d,1,mad)
d=d[rev(order(mads))[1:5000],]
@
If one wants to transform or normalize their data, they can easily do so using other Bioconductor methods or a simple statement.
We chose to use the default settings of the agglomerative heirarchical clustering algorithm using Pearson correlation distance, so it is appropriate to gene median center d using this simple statement:

<<>>=
d = sweep(d,1, apply(d,1,median,na.rm=T))
@

d is now ready for \texttt{ConsensusClusterPlus} analysis.

\subsection{Running \texttt{ConsensusClusterPlus}}
For this tutorial, we selected 80\% item resampling (pItem), 80\% gene resampling (pFeature),
a maximum evalulated k of 6 so that cluster counts of 2,3,4,5,6 are evaluated (maxK), 50 resamplings (reps), agglomerative heirarchical clustering algorithm (clusterAlg)
upon 1- Pearson correlation distances (distance), gave our output a title (title), and opted to have graphical results written to png files.  We also used a specific random seed so that this example is repeatable (seed).
\\
\\
** Note: In practice, a much higher reps is recommended such as 1,000 and a higher cluster count such as 20.


<<fig=FALSE>>=
library(ConsensusClusterPlus)
title=tempdir()
results = ConsensusClusterPlus(d,maxK=6,reps=50,pItem=0.8,pFeature=1,
title=title,clusterAlg="hc",distance="pearson",seed=1262118388.71279,plot="png")
@

<<echo=F,results=tex>>=
cat(sprintf("\\graphicspath{{%s}}", paste(gsub("[\\]","/",title),"/",sep="")))
cat("\n")
@

The output of \texttt{ConsensusClusterPlus} is a list, in which the element of the list corresponds to results from the \textit{kth} cluster, for instance, results[[2]] is the results
result of \textit{k}=2.  The seed option specifies a random number seed and is used here for reproducibility of this tutorial.  These list elements have the following elements:
<<echo=TRUE>>=
#consensusMatrix - the consensus matrix.  
#For .example, the top five rows and columns of results for k=2:
results[[2]][["consensusMatrix"]][1:5,1:5]

#consensusTree - hclust object 
results[[2]][["consensusTree"]]

#consensusClass - the sample classifications
results[[2]][["consensusClass"]][1:5]

#ml - consensus matrix result
#clrs - colors for cluster  
@

\subsection{Generating cluster and item consensus}
After executing \texttt{ConsensusClusterPlus}, one can optionally calculate cluster-consensus and item-consensus results by:

<<>>=
icl = calcICL(results,title=title,plot="png")
@

calcICL returns a list of two elements:
<<>>=
icl[["clusterConsensus"]]
@

<<>>=
icl[["itemConsensus"]][1:5,]
@


\section{Graphic Output Description}
The output of \texttt{ConsensusClusterPlus} consists of graphics, which are written to the screen, pdf file, or png files depending on the plot option; and numerical data 
which can be optionally written to a CSV file depending on the writeTable option.

\subsection{Consensus Matrices}
The first graphic shows the consensus color legend.
\\
<<echo=F,results=tex>>=
cat("\\includegraphics[width=60mm]{consensus001.png}",sep="")
@
\\
The remaining graphics are heatmaps of the consensus matrices for \textit{k} = 2, 3, 4, 5 \cite{monti}.  The consensus matrices have items as both rows and columns, which are microarrays in this .example, and where
consensus values range from 0 (never clustered together) to 1 (always clustered together) marked by white to dark blue.  The consensus matrices are
ordered by the consensus clustering which is depicted as a dendrogram atop the heatmap.  To aid analysis, the cluster memberships are marked by colored rectangles between
the dendrogram and heatmap according to a legend within the graphic.  This enables a user to compare a clusters' member count in the context of their
consensus.  
\\
<<echo=F,results=tex>>=
cat("\\includegraphics[width=60mm]{consensus002.png}",sep="")
cat("\\includegraphics[width=60mm]{consensus003.png}",sep="")
@
\\
<<echo=F,results=tex>>=
cat("\\includegraphics[width=60mm]{consensus004.png}",sep="")
cat("\\includegraphics[width=60mm]{consensus005.png}",sep="")
@
\\
\subsection{Consensus Cumulative Distribution Function (CDF) Plot}
This graphic shows the cumulative distribution functions \cite{monti} of the consensus matrix for each \textit{k} (indicated by colors), estimated by a histogram of 100 bins.  This figure allows a user to determine at what number of clusters, \textit{k},  the CDF reaches an approximate maximum, thus consensus and cluster confidence is at a maximum at this \textit{k}.  See \cite{monti} for further details intepretation.
\\
<<echo=F,results=tex>>=
cat("\\includegraphics[width=60mm]{consensus007.png}",sep="")
@
\\
\subsection{Delta Area Plot}
This graphic shows the relative change in area under the CDF curve \cite{monti} comparing \textit{k} and \textit{k} $-$ 1.  For \textit{k} = 2, there is no \textit{k} -1,
so the total area under the curve rather than the relative increase is plotted. This plot allows a user to determine the relative increase in consensus and determine \textit{k} at which there is no appreciable increase.  See \cite{monti} for intepretation.
\\ 
<<echo=F,results=tex>>=
cat("\\includegraphics[width=60mm]{consensus008.png}",sep="")
@
\\
\subsection{Tracking Plot}
This graphic shows the cluster assignment of items (columns) for each \textit{k} (rows) by color.  
The colors correspond to the colors of the consensus matrix class asssignments.  Hatch marks below the plot indicate items/samples.
This plot provides a view of item cluster membership across different \textit{k} and enables a user to track the history of clusters relative to earlier clusters.
Items that change clusters often (changing colors within a column) are indicative of unstable membership.  Clusters with an abundance of unstable members
suggest an unstable cluster.
\\
<<echo=F,results=tex>>=
cat("\\includegraphics[width=60mm]{consensus009.png}",sep="")
@
\\
\subsection{Cluster-Consensus Plot}
This graphic shows the \textit{cluster-consensus} value of clusters at each \textit{k}.  This is the mean of all pairwise consensus values
between a cluster's members.  Cluster is indicated by color following the same color scheme as the cluster matrices and tracking plots.
The bars are grouped by \textit{k} which is marked on the horizontal axis.
High values indicate a cluster has high stability and low values indicate a cluster has low stability.  This plot enables a user to view the
mean cluster-consensus among clusters at a given \textit{k} and compare values of clusters across different \textit{k} via the color scheme.
\\
<<echo=F,results=tex>>=
cat("\\includegraphics[width=60mm]{icl003.png}",sep="")
@
\\
\subsection{Item-Consensus Plot}
\textit{Item-consensus} values are the mean consensus of an item with all items in a particular cluster.  An item has \textit{k} item-consensus values corresponding to each cluster
at a particular \textit{k}.  These values are depicted in barplots for each \textit{k}.  Samples are stacked bars.  \textit{Item-consensus} values are indicated by the heights of the colored portion of 
the bars, whose color corresponds to the common color scheme.  Bars' rectangles are ordered
by increasing value from bottom to top.  The asterisks at the top indicate the consensus cluster for each item.
\\
<<echo=F,results=tex>>=
cat("\\includegraphics[width=60mm]{icl001.png}",sep="")
@
\\
This plot provides a view of item-consensus across all other clusters at a given \textit{k}.  This enables a user to see if a sample is a very "pure" member of a
cluster or if it shares high consensus to multiple clusters (large rectangles in a column of multiple colors), suggesting that it is an unstable or "unpure" member.  These values could be used to
select "core" samples similar to \cite{verhaak} that are highly representative of a cluster.  Further, this plot can aid cluster number decisions.  For instance, if a cluster
consists mainly of members with very "unpure" items, then this evidence could be used to support a maximum cluster number at 1 below this \textit{k} or this evidence
could support that this cluster is an outlier cluster.  Decisions such as these are best to be made by the user in conjunction with other evidence such as consensus matrices, tracking plots, etc.
\\
\\
\subsection{Additional options for \texttt{ConsensusClusterPlus} function}

\begin{itemize}

\item \textbf{clusterAlg} This option specifies the type of clustering algorithm to use: "hc" for heirarchical clustering, and "km" for kmeans.  Alternatively, one can
supply their own clustering hook, which should accept a distance matrix and a cluster number as its arguments and returns vector of cluster assignments having the same order as the distance matrix columns.
For .example, this simple function executes divisive clustering using the diana function from the cluster package and returns the expected object.  The last line shows an .example of how this could be used.

<<>>=
#library(cluster)
#dianaHook = function(this_dist,k){
  #tmp = diana(this_dist,diss=TRUE)
  #assignment = cutree(tmp,k)
  #return(assignment)  
#}
#ConsensusClusterPlus(d,clusterAlg="dianaHook",distance="pearson",...)
TRUE
@

\item \textbf{innerLinkage} This option specifies the linkage method to use in iterative agglomerative heirarchical clustering.  Not applicable to other cluster algorithms.
\item \textbf{finalLinkage} This option specifies the linkage method to use in the final agglomerative heirarchical clustering.
\item \textbf{distance} This option specifies the distance metric to use: "pearson" for 1-Pearson correlation coefficient, "spearman" for 1-Spearman correlation coefficient, "euclidean" for Euclidean distance.
Note: pearson is suggested for heirarchical clustering and euclidean for kmeans.
\item \textbf{tmyPal} character vector of ordered colors to use for consensus matrix.  If not specified, a series of white to blue colors is used.
\item \textbf{writeTable} boolean. If TRUE, write consensus matrices, ICL, and log to file.  
\item \textbf{weightsFeature} numerical vector of weights for sampling features.  See help(sample) for further details. 
\item \textbf{weightsItem} numerical vector of weights for sampling items.  See help(sample) for further details.
\item \textbf{verbose} boolean.  If TRUE, print messages to the screen to indicate progress.  This is useful for large datasets.

\end{itemize}


\begin{thebibliography}{}
\bibitem{monti}Monti, S., Tamayo, P., Mesirov, J., Golub, T. (2003) Consensus Clustering: A Resampling-Based Method for Class Discovery and Visualization of Gene Expression Microarray Data. Machine Learning, 52, 91$-$118.
\bibitem{hayes}Hayes, D.N, Monti, S., Parmigiani, G. et al. (2006) Gene Expression Profiling Reveals Reproducible Human Lung Adenocarcinoma Subtypes in Multiple Independent Patient Cohorts. Journal of Clinical Oncology, 24 (31) 5079$-$5090.
\bibitem{verhaak}Verhaak, R., Hoadley, K., et al. (2010) Integrated genomic analysis identifies clinically relevant subtypes of glioblastoma characterized by abnormalities in PDGFRA, IDH1, EGFR and NF1. Cancer Cell. 17,1-13.
\bibitem{gp}\texttt{http://www.broadinstitute.org/cancer/software/genepattern/}

\end{thebibliography} 

\subsection{Changes}
\begin{itemize}

\item Version 1.0.1.  Item-consensus calculation was corrected.  Consensus matric heat maps are now guaranteed to correspond to the scale.
\item Version 1.5.1.  Version 1.0.1 changes were re-incorporated into Bioc 2.9, 2.8.  Version 1.0.1 was part of Bioc 2.6, but not part of Bioc 2.7.  

\end{itemize}


\end{document}