Copy number routines in the \crlmm{} package are available for
  Affymetrix 5.0 and 6.0 platforms, as well as several Illumina
  platforms.  This vignette assumes that the arrays have already been
  successfully preprocessed and genotyped as per the instructions in
  the \verb+AffyGW+ and \verb+IlluminaPreprocessCN+
  vignettes for the Affymetrix and Illumina platforms,
  respectively. While this vignette uses Affymetrix 6.0 arrays for
  illustration, the steps at this point are identical for both
  platforms.  See \citep{Scharpf2011} for details regarding the
  methodology implemented in \crlmm{} for copy number analysis.  In
  addition, a compendium describing copy number analysis using the
  \crlmm{} package is available from the author's website:

\textbf{Limitations:} While a minimum number of samples is not
required for preprocessing and genotyping, copy number estimation in
the \crlmm{} package currently requires at least 10 samples per batch.
The parameter estimates for copy number and the corresponding
estimates of raw copy number will tend to be more noisy for batches
with small sample sizes (e.g., $<$ 50).  Chemistry plate or scan date
are often useful surrogates for batch.  Samples that were processed at
similar times (e.g., in the same month) can be grouped together in the
same batch.

\section{Quality control}

The signal to noise ratio (SNR) estimated by the CRLMM genotyping
algorithm is an overall measure of the separation of the diallelic
genotype clusters at polymorphic loci and can be a useful measure of
array quality.  Small SNR values can indicate possible problems with
the DNA.  Depending on the size of the dataset and the number of
samples with low SNR, users may wish to rerun the preprocessing and
genotyping steps after excluding samples with low SNR.  The SNR is
stored in the \verb+phenoData+ slot of the \Robject{CNSet} object and
is available after preprocessing and genotyping. SNR values below 5
for Affymetrix or below 25 for Illumina may indicate poor sample
quality.  The following code chunk makes a histogram of the SNR values
for the HapMap samples.

<<snr,fig=TRUE,include=FALSE,width=6, height=4>>=
snr <- cnSet$SNR[]
		breaks=25, xlim=range(snr), xlab="SNR"))

\section{Copy number estimation}

There are two ways to obtain marker-level estimates of copy number
that are supported by \Rpackage{crlmm}. One approach is to fit a
linear model to the normalized intensities stratified by the diallelic
genotype call at each SNP, as described in \cite{Scharpf2011}.
Another alternative is to compute log R ratio and B allele
frequencies. The latter is often better supported by downstream hidden
Markov models such as those in the \Rpackage{VanillaICE} package. We
describe each approach in the following two sections.

\subsection{Linear model for normalized intensities}

In this section, we fit a linear model to the normalized intensities
stratified by the diallic genotype call.  The intercept and slope from
the linear model are both SNP- and batch-specific.  The implementation
in the \crlmm{} package is encapsulated by the function
\Rfunction{crlmmCopynumber} that, using the default settings, can be
called by passing a single object of class \Rclass{CNSet}.


The following steps were performed by the \Rfunction{crlmmCopynumber}
\item sufficient statistics for the genotype clusters for
  each batch
\item unobserved genotype centers imputed
\item posterior summaries of sufficient statistics
\item intercept and slope for linear model
Depending on the value of \verb+ocProbesets()+, these summaries are
computed for subsets of the markers to reduce the required RAM. Note
that the value returned by the \Rfunction{crlmmCopynumber} function in
the above example is \verb+TRUE+.  The reason the function returns
\verb+TRUE+ in the above example is that the elements of the
\verb+batchStatistics+ slot have the class \Rclass{ff\_matrix}.
Rather than keep the statistical summaries in memory, the summaries
are written to files on disk using protocols described in the
\Rpackage{ff} package. Hence, while the \Robject{cnSet} object itself
is unchanged as a result of the \Rfunction{crlmmCopynumber} function,
the data on disk is updated accordingly.  Users that are interested in
accessing these low-level summaries can refer to the
\verb+Infrastructure+ vignette.  Note that the data structure depends
on whether the elements of the \verb+batchStatistics+ slot are
\Robject{ff} objects or ordinary matrices.  In this example, the
elements of \verb+batchStatistics+ have the class \Rclass{ff\_matrix}.

nms <- ls(batchStatistics(cnSet))
cls <- rep(NA, length(nms))
for(i in seq_along(nms)) cls[i] <- class(batchStatistics(cnSet)[[nms[i]]])[1]
all(cls == "ff_matrix")

The batch-specific statistical summaries computed by
\Robject{crlmmCopynumber} are written to files on disk using protocols
described in the \R{} package \Rpackage{ff}. The value returned by
\Rfunction{crlmmCopynumber} is \verb+TRUE+, indicating that the files
on disk have been successfully updated.  Note that while the
\Robject{cnSet} object is unchanged, the values on disk are different.

chr1.index <- which(chromosome(cnSet) == 1)

cnSet2 <- cnSet[chr1.index, ]

for(i in seq_along(nms)) cls[i] <- class(batchStatistics(cnSet2)[[nms[i]]])[1]
all(cls == "matrix")

<<clean, echo=FALSE, results=hide>>=
rm(cnSet2); gc()

\paragraph{Raw total copy number.}
Several functions are available that will compute relatively quickly
the allele-specific, \emph{raw} copy number estimates. At allele $k$,
marker $i$, sample $j$, and batch $p$, the estimate of allele-specific
copy number is computed by subtracting the estimated background from
the normalized intensity and scaling by the slope coefficient. More
formally, \newcommand{\A}{A} \newcommand{\B}{B}
{\hat c}_{k,ijp} &=& \mbox{max}\left\{\frac{1}{{\hat
    \phi}_{k,ip}}\left(I_{k,ijp}-{\hat \nu}_{k,ip}\right), ~0\right\}
\mbox{~for~} k \in \{\A, \B\}.
\noindent See \cite{Scharpf2011} for details.

The function \Rfunction{totalCopynumber} translates the normalized
intensities to an estimate of raw copy number by adding the
allele-specific summaries in Equation \eqref{eq:cnK}. For large
datasets, the calculation will not be instantaneous as the I/O can be
substantial.  Users should specify either a subset of the markers or a
subset of the samples to avoid using all of the available RAM.  For
example, in the following code chunk we compute the total copy number
at all markers for the first 2 samples, and the total copy number for
chromosome 20 for the first 50 samples.

tmp <- totalCopynumber(cnSet, i=seq_len(nrow(cnSet)), j=1:2)
tmp2 <- totalCopynumber(cnSet, i=which(chromosome(cnSet) == 20), j=seq_len(ncol(cnSet)))

Alternatively, the functions \Rfunction{CA} and \Rfunction{CB} compute
the allele-specific copy number.  For instance, the following code
chunk computes the allele-specific summaries at all polymorphic loci
for the first 2 samples.

snp.index <- which(isSnp(cnSet) & !
ca <- CA(cnSet, i=snp.index, j=1:2)
cb <- CB(cnSet, i=snp.index, j=1:2)

\subsection{Container for log R ratios and B allele frequencies}

A useful container for storing the \crlmm{} genotypes, genotype
confidence scores, and the total or relative copy number at each
marker is the \Rclass{BafLrrSetList} class.  Coercion of a
\Rclass{CNSet} object to a \Rfunction{BafLrrSetList} object can be
acheived by the function \Rfunction{BafLrrSetList} as
illustrated below. Users should note that if the \verb+assayData+
elements in the \Rclass{CNSet} instance are \Rpackage{ff} objects, the
\verb+assayData+ elements of each element in the \Rclass{BafLrrSetList}
object will be \Rpackage{ff}-dervied objects (a new
\verb+total_cn*.ff+ file will be created in the \verb+ldPath()+
directory). The following code-chunk is not evalutated.

Estimation of log R ratios and B allele frequencies from an object of class \Rclass{CNSet} instantiated from genotyping Affymetrix CEL files or Illumina iDat files requires running the \R{} function \Rfunction{crlmmCopynumber} as a preliminary step.  Specifying \texttt{fit.linearModel=FALSE} will speed up computation as this function will by default compute summary statistics unnecessary for computing BAFs and log R ratios.

crlmmCopynumber(cnSet, fit.linearModel=FALSE)

oligoSetList <- BafLrrSetList(cnSet)
## oligoSnpSet of first chromosome

\noindent Log R ratios and B allele frequences can be retrieved by the
accessors \Rfunction{lrr} and \Rfunction{baf}, respectively.

lrrList <- lrr(oligoSetList)
dim(lrrList[[1]]) ## log R ratios for chromosome 1.
bafList <- baf(oligoSetList)
dim(bafList[[1]]) ## B allele frequencies for chromosome 1

If the \Rfunction{crlmmCopynumber} function is not run prior to the
\Rfunction{BafLrrSetList} construction, the log R ratios and BAFs will
be initialized as NAs.