%
% NOTE -- ONLY EDIT THE .Rnw FILE!!!  The .tex file is
% likely to be overwritten.
%
%\VignetteIndexEntry{Notes for eSet developers}
%\VignetteDepends{}
%\VignetteKeywords{Expression Analysis}
%\VignettePackage{Biobase}
\documentclass[]{article}

\usepackage{times}
\usepackage{hyperref}

\newcommand{\Rfunction}[1]{{\texttt{#1}}}
\newcommand{\Robject}[1]{{\texttt{#1}}}
\newcommand{\Rpackage}[1]{{\textit{#1}}}
\newcommand{\Rmethod}[1]{{\texttt{#1}}}
\newcommand{\Rfunarg}[1]{{\texttt{#1}}}
\newcommand{\Rclass}[1]{{\textit{#1}}}

\textwidth=6.2in
\textheight=8.5in
%\parskip=.3cm
\oddsidemargin=.1in
\evensidemargin=.1in
\headheight=-.3in

\newcommand{\R}{\textsf{R}}
\newcommand{\eSet}{\Robject{eSet}}
\newcommand{\ExpressionSet}{\Robject{ExpressionSet}}
\newcommand{\Biobase}{\Rpackage{Biobase}}

\newcommand{\term}[1]{\emph{#1}}
\newcommand{\mref}[2]{\htmladdnormallinkfoot{#2}{#1}}

\begin{document}

\title{\Biobase{} development and the new \eSet}
\author{\mref{mailto:mtmorgan@fhcrc.org}{Martin T. Morgan}}
\date{7 August, 2006}
\maketitle

Revised 4 September, 2006 -- \Robject{featureData} slot.
Revised 20 April 2007 -- minor wording changes; \Rfunarg{verbose} and
other arguments passed through \Rmethod{updateObject} example;
introduce a second variant of \Rmethod{initialize} illustrating its
use as a copy constructor.
Revised 4 November, 2010 -- \Robject{experimentData} slot as \Rclass{MIAxE} 
class. \Rclass{MIAME} class is now subclass of \Rclass{MIAxE}.

\section{Introduction}

These notes help \emph{developers} who are interested in using and
extending the \eSet{} class hierarchy, and using features in
\Biobase. The information here is not useful to regular users of
\Biobase.

This document illustrates the \Biobase{} structures and approaches
that make it it easy for developers to creatively use and extend the
\eSet{} class hierarchy.

The document starts with a brief description of the motivation for
change, and a comparison of the old (before August, 2006) and new
\eSet{}s and related functionality (e.g.,the \Rclass{Versioned} class
and \Rfunction{updateObject} methods). We then illustrate how \eSet{}
can be extended to handle additional types of data, and how new
methods can exploit the \eSet{} class hierarchy. We conclude with a
brief summary of lessons learned, useful developer-related
side-effects of efforts to revise \eSet{}, and possible directions for
future development.

\section{Comparing old and new}

What is an \eSet{}?
\begin{itemize}
\item Coordinate high through-put (e.g., gene expression) and phenotype data.
\item Provide common data container for diverse Bioconductor packages.
\end{itemize}
Motivation for change (August, 2006).
\begin{itemize}
\item What was broken?  Complex data structure. Inconsistent object
  validity. No straight-forward way to extend \eSet{} to new data
  types. 
\item What forward-looking design goals did we have? Flexible storage
  model. Class hierarchy to promote code reuse and facilitate
  extension to new data objects. Methods for updating serialized instances.
\end{itemize}
Key features in the redesign.
\begin{itemize}
\item Simplified data content.
\item Structured class hierarchy .
\item Alternative storage modes.
\item More validity checking.
\item Conversion of example data in \Biobase{}, and many other data
  sets elsewhere in Bioconductor, to \Rclass{ExpressionSet}.
\item \Rclass{Versioned} class information (in the development branch).
\item \Rfunction{updateObject} methods (in the development branch).
\end{itemize}

\section{A quick tour}
<<Biobase,echo=FALSE,results=hide>>=
options(width=69)
library(Biobase)
@ 

\subsection{The \eSet{} object: high-throughput experiments}
Purpose.
\begin{itemize}
\item Coordinate and contain high-throughput genomic data.
\end{itemize}
Structure: virtual base class.
<<eSet-class>>=
getClass("eSet")
@ 
\begin{itemize}
\item \Robject{assayData}: high-throughput data.
\item \Robject{phenoData}: sample covariates.
\item \Robject{featureData}: feature covariates.
\item \Robject{experimentData}: experimental description.
\item \Robject{annotation}: assay description.
\item See below, and \verb|?"eSet-class"|
\end{itemize}

\subsubsection{\Robject{assayData}: high-throughput data}
Purpose.
\begin{itemize}
\item Efficiently and flexibly contain high-volume data.
\end{itemize}
Structure: \Rclass{list}, \Rclass{environment}, or \Rclass{lockEnvironment} class union.
\begin{itemize}
\item Each element of \Rclass{list} / \Rclass{environment} / \Rclass{lockEnvironment} is a matrix
\item Rows: \emph{features}, e.g., gene names.
\item Columns: \emph{samples} represented on each chip. 
\item All matricies must have the same dimensions, row names, and column names.
\item Subclasses determine which matricies \emph{must} be present.
\item See \verb|?"AssayData-class"|
\end{itemize}

\subsubsection{\Robject{phenoData}: sample covariates}
Purpose.
\begin{itemize}
\item Contain and document sample covariates.
\end{itemize}
Structure: \Rclass{AnnotatedDataFrame}.
\begin{itemize}
\item \Robject{data}: \Rclass{data.frame}.
  \begin{itemize}
  \item Rows: sample identifiers.
  \item Columns: measured covariates.
  \end{itemize}
\item \Robject{varMetadata}: \Rclass{data.frame}.
  \begin{itemize}
  \item Rows: measured covariate labels.
  \item Columns: covariate descriptors. 
  \end{itemize}
\item See \verb|?"AnnotatedDataFrame-class"|
\end{itemize}

\subsubsection{\Robject{featureData}: feature covariates}
Purpose.
\begin{itemize}
\item Contain and document feature covariates specific to the
  experiment; use the \Robject{annotation} slot for chip-level descriptions.
\end{itemize}
Structure: \Rclass{AnnotatedDataFrame}.
\begin{itemize}
\item \Robject{data}: \Rclass{data.frame}.
  \begin{itemize}
  \item Rows: feature identifiers. These match row names of \Robject{assayData}.
  \item Columns: measured covariates.
  \end{itemize}
\item \Robject{varMetadata}: \Rclass{data.frame}.
  \begin{itemize}
  \item Rows: measured covariate labels.
  \item Columns: covariate descriptors. 
  \end{itemize}
\item See \verb|?"AnnotatedDataFrame-class"|
\end{itemize}

\subsubsection{\Robject{experimentData}: experiment description}
Purpose.
\begin{itemize}
\item Summarize where and how the experiment was performed.
\end{itemize}
Structure: \Rclass{MIAxE} virtual class
\begin{itemize}
\item[] {In Biobase version 2.11 the \Rclass{MIAxE} virtual class was introduced 
as a general container for experiment meta-data. The \Rclass{MIAME} class
is now a subclass of \Rclass{MIAxE}. See \verb|?"MIAxE-class"|. 
The \Rclass{MIAME} class structure is as follows} 
\end{itemize}
\begin{itemize}
\item \Robject{title}: experiment title.
\item \Robject{name}: experimenter name(s).
\item \Robject{preprocessing}: \Robject{list} of pre-processing steps.
\item Additional slots.
\item See \verb|?"MIAME-class"|.
\end{itemize}

\subsubsection{\Robject{annotation}: assay description}
Purpose.
\begin{itemize}
\item Link experiment to annotation package.
\end{itemize}
Structure: \Rclass{character}
\begin{itemize}
\item Label identifying annotation package.
\end{itemize}

%% Methods

\subsection{Important \eSet{} methods}
Initialization.
\begin{itemize}
\item \eSet{} is \verb|VIRTUAL|, initialize via subclass \Rfunction{callNextMethod}
\end{itemize}
Accessors (get, set).
\begin{itemize}
\item \Rfunction{assayData(obj); assayData(obj) <- value}: access or assign \Robject{assayData}
\item \Rfunction{phenoData(obj); phenoData(obj) <- value}: access or assign \Robject{phenoData}
\item \Rfunction{experimentData(obj); experimentData(obj) <- value}: access or assign \Robject{experimentData}
\item \Rfunction{annotation(obj); annotation(obj) <- value}: access or assign \Robject{annotation}
\end{itemize}
Subsetting.
\begin{itemize}
\item \Rfunction{obj[i, j]}: select genes $i$ and samples $j$.
\item \Rfunction{obj\$name; obj\$name <- value}:
  retrieve  or assign covariate \Robject{name} in \Robject{phenoData}
\end{itemize}

\subsubsection{Additional \eSet{} methods}
\begin{itemize}
\item \Rfunction{show}.
\item \Rfunction{storageMode}: influence how \Robject{assayData} is stored.
\item \Rfunction{updateObject}: update \eSet{} objects to their current version.
\item \Rfunction{validObject}: ensure that \eSet{} is valid.
\end{itemize}

The \Rfunction{validObject} method is particularly important to
\eSet{}, ensuring that \eSet{} contains consistent structure to data.
<<eSet-validity>>=
getValidity(getClass("eSet"))
@ 
The validity methods for \eSet{} reflect our design goals. All
\Robject{assayData} members must have identical row and column
dimensions and \Robject{featureNames}. The names and numbers of
samples must be the same in \Robject{assayData} and
\Robject{phenoData} slots. Validity methods are defined for the
classes underlying each slot as well. For instance, the validity
methods for \Robject{AnnotatedDataFrame} check that variables used in
\Robject{pData} are at least minimally described in \Robject{varMetadata}.

%% Subclasses

\subsection{Subclasses of \eSet}
\Biobase{} defines three classes that extend \eSet{}. \ExpressionSet{}
(discussed further below) is meant to contain microarray gene
expression data. \Rclass{SnpSet} is a preliminary class to contain SNP
data; other classes in development (e.g., in \Rpackage{oligo}) may
provide alternative implementations for SNP data. \Rclass{MultiSet} is
an \ExpressionSet-like class, but without restriction on the names
(though not structure) of elements in the \Robject{assayData} slot.

\subsubsection{\ExpressionSet}
Purpose: 
\begin{itemize}
\item Contain gene expression data.
\end{itemize}
Required \Robject{assayData} members.
\begin{itemize}
\item \Robject{exprs}, a matrix of expression values.
\end{itemize}
Important methods.
\begin{itemize}
\item Initialization (additional details below):
<<ExpressionSet-initialize,eval=FALSE>>=
obj <- new("ExpressionSet", phenoData = new("AnnotatedDataFrame"), experimentData = new("MIAME"), annotation = character(), exprs = new("matrix")) 
@ 
\item \Rfunction{exprs(obj), exprs(obj) <- value}: get or set \Robject{exprs}; methods defined for \Rclass{ExpressionSet}, \Rclass{AssayData}.
\end{itemize}
 

\subsubsection{\Rclass{MultiSet} and \Rclass{SnpSet}}
\Rclass{MultiSet}.
\begin{itemize}
\item Purpose: flexibly contain a collection of expression data matricies.
\item Required \Robject{assayData} members: none.
\end{itemize}
\Rclass{SnpSet}.
\begin{itemize}
\item Purpose: contain genomic SNP calls.
\item Required \Robject{assayData} members: \Robject{call}, \Robject{callProbability}.
\end{itemize}

\section{Comments on \Robject{assayData}: high-throughput data storage}
The \Robject{assayData} slot is meant to store high-throughput
data. The idea is that the slot contains identically sized matrices
containing expression or other data. All matrices in the slot must
have the same dimension, and are structured so that rows represent
`features' and columns represent `samples'. Validity methods enforce
that row and column names of slot elements are identical. 

For technical reasons, creating instance of \Rclass{AssayData} is
slightly different from the way this is usually done in
\R{}. Normally, one creates an instance of a class with an expression
like \Rfunction{new("ExpressionSet", ...)}, with the \ldots
representing additional arguments. \Rclass{AssayData} objects are
created with
<<newAssayData,eval=FALSE>>=
assayDataNew("environment", elt)
@ 
where \Robject{elt} might be a matrix of expression values. For the
curious, the reason for this setup stems from our desire to have a
class that \Rfunction{is} a list or environment, rather than a class
that has a slot that contains a list or environment. The
\Rfunction{is} relationship is desirable to avoid unnecessary
function calls to access slots, and requires that a class
\Rfunction{contain} the base type (e.g., \Rclass{environment}). Until
recently an \R{} object could not \Rfunction{contain} an
\Rclass{environment}.

The \Robject{assayData} slot of \ExpressionSet{} objects must contain
a matrix named \Robject{exprs}. Nonetheless, the \ExpressionSet{}
validity method tries to be liberal -- it guarantees that the object
has an \Robject{exprs} element, but allows for other elements too. The
prudent developer wanting consistent additional data elements should
derive a class from \ExpressionSet{} that enforces the presence of
their desired elements.

The \Rclass{AssayData} class allows for data elements to be stored in
three different ways (see \Rfunction{?storageMode} and
\Rfunction{?"storageMode<-"} for details): as a \Robject{list},
\Robject{environment}, or \Robject{lockedEnvironment}.  Developers are
probably familiar with \Robject{list} objects; a drawback is that
\Robject{exprs} elements may be large, and some operations on lists in
\R{} may trigger creation of many copies of the the \Robject{exprs}
element. This can be expensive in both space and time. Environments
are nearly unique in \R{}, in that they are passed by reference rather
than value. This eliminates some copying, but has the unfortunate
consequence that side-effects occur -- modifications to an environment
inside a function influence the value of elements outside the
function. For these reasons, environments can be useful as `read only'
arguments to functions, but can have unexpected consequences when
functions modify their arguments.  Locked environments implemented in
\Biobase{} try to strike a happy medium, allowing pass by reference for
most operations but triggering (whole-environment) copying when
elements in the environment are modified. The locking mechanism is
enforced by only allowing known `safe' operations to occur, usually by
channeling user actions through the accessor methods:
<<assayData-storageMode>>=
data(sample.ExpressionSet)
storageMode(sample.ExpressionSet)
tryCatch(assayData(sample.ExpressionSet)$exprs <- log(exprs(sample.ExpressionSet)),
         error=function(err) cat(conditionMessage(err)))
exprs(sample.ExpressionSet) <- log(exprs(sample.ExpressionSet))
@ 
The \Rfunction{setReplaceMethod} for \Robject{exprs} (and
\Robject{assayData}) succeeds by performing a deep copy of the entire
environment. Because this is very inefficient, the recommended paradigm
to update an element in a \Robject{lockedEnvironment} is to extract
it, make many changes, and then reassign it. Developers can study
\Robject{assayData} methods to learn more about how to lock and unlock
environment bindings.  \Biobase{} allows the experienced user to
employ (and run the risks of) environments, but the expectation is
that most user objects are construced with the default
\Robject{lockedEnvironment} or \Robject{list}.

A longer term consideration in designing \Rclass{AssayData} was to
allow more flexible methods of data storage, e.g., through
database-hosted arrays. This is facilitated by using generic functions
such as \Rfunction{exprs()} for data access, so that classes derived
from \Rclass{AssayData} can provide implementations appropriate for
their underlying storage mode.


\section{Extending \eSet}
A designer wanting to implement \eSet{} for a particular type of data
creates a class that `contains' \eSet{}. The steps for doing this are
described below. One example of such a class is \ExpressionSet,
designed to hold a matrix of gene expression values in the
\Robject{assayData} slot. 
<<ExpressionSet-class>>=
getClass("ExpressionSet")
getValidity(getClass("ExpressionSet"))
@ 
The data structure of an \ExpressionSet{} is identical to that of
\eSet, and in fact is inherited (without additional slot creation)
from \eSet{}. The main difference is that the validity methods of
\eSet{} are augmented by a method to check that the
\Robject{assayData} slot contains an entity named \Robject{"exprs"}. A
valid \ExpressionSet{} object must also satisfy all the validity
requirements of \eSet{}, but the developer does not explicitly invoke
validity checking of the parts of the data structure inherited from
\eSet{}.

\subsection{Implementing a new class: a \Rclass{SwirlSet} example}

We want the \emph{Swirl} data set (see the \Robject{SW} two color data
set that motivates this class) to contain four elements in the
\Robject{assayData} slot: \Robject{R}, \Robject{G}, \Robject{Rb},
\Robject{Gb}. To derive a class from \eSet{} for this data, we create
a class, and provide initializaation and validation methods.

We create a class as follows:
<<SwirlSet-class>>=
setClass("SwirlSet", contains="eSet")
@ 
Notice that there are no new data elements in \Rclass{SwirlSet}
compared with \eSet{}. The \Rfunction{initialize} method is written as
<<SwirlSet-initialize>>=
setMethod("initialize", "SwirlSet",
          function(.Object,
                   R = new("matrix"),
                   G = new("matrix"),
                   Rb = new("matrix"),
                   Gb = new("matrix"),
                   ...) {
            callNextMethod(.Object,
                           R=R, G=G, Rb=Rb, Gb=Gb,
                           ...)
        })
@ 
% 
A slightly different \Rmethod{initialize} method allows the user to
specify either the \Robject{assayData} or the \Robject{assayData}
content. In advanced use, this has the advantage that
\Rmethod{initialize} can be used as a `copy constructor' to update
several slots simultaneously.
<<SwirlSet-initialize-2>>=
setMethod("initialize", "SwirlSet",
          function(.Object,
                   assayData=assayDataNew(
                     R=R, G=G, Rb=Rb, Gb=Gb),
                   R = new("matrix"),
                   G = new("matrix"),
                   Rb = new("matrix"),
                   Gb = new("matrix"),
                   ...) {
            if (!missing(assayData) && 
                any(!missing(R), !missing(G), !missing(Rb), !missing(Gb))) {
                warning("using 'assayData'; ignoring 'R', 'G', 'Rb', 'Gb'")
            }
            callNextMethod(.Object, assayData=assayData, ...)
        })
@ 

The structure of the \Rfunction{initialize} method is a bit different
from those often seen in \R{}. Often, \Rfunction{initialize} has only
\Robject{.Object} as a named argument, or, if there are other named
arguments, they correspond to slot names. Here our initialize method
accepts four arguments, named after the \Robject{assayData} elements.
Inside the \Rfunction{initialize} method, the named arguments are
passed to the next initialization method in the hierarch (i.e.,
\Rfunction{initialize} defined for \eSet). The \eSet{}
\Rfunction{initialize} method then uses these arguments to populate
the data slots in \Robject{.Object}. In particular, \eSet{} places all
arguments other \Rfunction{phenoData}, \Rfunction{experimentData}, and
\Rfunction{annotation} into the \Robject{assayData} slot.  The \eSet{}
\Rfunction{initialize} method then returns the result to the
\Rfunction{initialize} method of \Rclass{SwirlSet}, which returns a
\Rclass{SwirlSet} object to the user:
<<SwirlSet-new>>=
new("SwirlSet")
@ 

General programing guidelines emerge from experience with the
\Rfunction{initialize} method of \eSet{} and derived classes. First,
an appropriate strategy is to name only those data elements that will
be manipulated directly by the \Rfunction{initialize} method. For
instance, the definition above did \emph{not} name \Robject{phenoData}
and other \eSet{} slots by name. To do so is not incorrect, but would
require that they be explicitly named (e.g.,
\Rfunction{phenoData=phenoData}) in the \Rfunction{callNextMethod}
code. Second, the arguments \Robject{R, G, Rb, Rg} are present in the
\Rfunction{initialize} method to provide defaults consistent with
object construction; the `full' form of \Rfunction{callNextMethod},
replicating the named arguments, is required in the version of \R{} in
which this class was developed. Third, named arguments can be
manipulated before \Rfunction{callNextMethod} is invoked.  Fourth, the
return value of \Rfunction{callNextMethod} can be captured\ldots
<<initialize-.Object,eval=FALSE>>=
setMethod("initialize", "MySet",
          function(.Object, ...) {
              .Object <- callNextMethod(.Object, ...)
          })
              .
@ 
and manipulated before being returned to the user. Finally, it is the
responsbilitiy of the developer to ensure that a valid object is
created; \Rfunction{callNextMethod} is a useful way to exploit
correctly designed \Rfunction{initialize} methods for classes that the
object extends, but the developer is free to use other techniques to
create valid versions of their class.

A validity method might complete our new class. A validity method is
essential to ensure that the unique features of \Rclass{SwirlSet} --
our reason for designing the new class -- are indeed present. We
define our validity method to ensure that the \Robject{assayData} slot
contains our four types of expression elements:
<<SwirlSet-validity>>=
setValidity("SwirlSet", function(object) {
  assayDataValidMembers(assayData(object), c("R", "G", "Rb", "Gb"))
})
@ 

Notice that we do not have to explicitly request that the validity of
other parts of the \Rclass{SwirlSet} object are valid; this is done
for us automatically. Objects are checked for validity when they are
created, but not when modified. This is partly for efficiency reasons,
and partly because object updates might transiently make them
invalid. So a good programing practice is to ensure validity after
modification, e.g.,
<<validity-sometimes,eval=FALSE>>=
myFancyFunction <- function(obj) {
  assayData(obj) <- fancyAssaydData # obj invalid...
  phenoData(obj) <- justAsFancyPhenoData # but now valid
  validObject(obj)
  obj
}
@ 
Assigning \Robject{fancyAssaydData} might invalidate the object, but
\Robject{justAsFancyPhenoData} restores validity.

\section{\Rclass{Versioned}}

One problem encountered in the Bioconductor project is that data
objects stored to disk become invalid as the underlying class
definition changes. For instance, earlier releases of \Biobase{}
contain a sample \Rclass{eSet} object. But under the changes discussed
here, \eSet{} is virtual and the stored object is no longer valid. The
challenge is to easily identify invalid objects, and to provide a
mechanism for updating old objects to their new representation.

\Biobase{} introduces the \Rclass{Versioned} and
\Rclass{VersionedBiobase} classes to facilitate this. These classes
are incorporated into key \Biobase class definitions.\Biobase also
defines the \Rfunction{updateObject} generic and methods for
conveniently updating old objects to their new representation.
<<updateObject-eg>>=
data(sample.ExpressionSet)
classVersion(sample.ExpressionSet)
obj <- updateObject(sample.ExpressionSet)
@ 
The version information for this object is a named list. The first two
elements indicate the version of R and Biobase used to create the
object. The latter two elements are contained in the class prototype,
and the class prototype is consulted to see if the instance of an
object is 'current'. These lists can be subsetted in the usual way,
e.g.,
<<isCurrent>>=
isCurrent(sample.ExpressionSet)[c("eSet", "ExpressionSet")]
@ 

Versioned classes, \Rfunction{updateObject} and related methods
simplify the long-term maintenance of data objects. Take the fictious
\Rclass{MySet} as an example. 
<<MultiSet-obj>>=
setClass("MySet",
         contains = "eSet",
         prototype = prototype(
           new("VersionedBiobase",
               versions=c(classVersion("eSet"), MySet="1.0.0"))))
obj <- new("MySet")
classVersion(obj)
@ 
This is a new class, and might undergo changes in its structure at
some point in the future. When these changes are introduced, the
developer will change the version number of the class in its prototype
(the last line, below):
<<MultiSetRevised>>=
setClass("MySet",
         contains = "eSet",
         prototype = prototype(
           new("VersionedBiobase",
               versions=c(classVersion("eSet"), MySet="1.0.1"))))
isCurrent(obj)
@ 
and add code to update to the new version
<<updateObject-MultiSet>>=
setMethod("updateObject", signature(object="MySet"),
          function(object, ..., verbose=FALSE) {
              if (verbose) message("updateObject(object = 'MySet')")
              object <- callNextMethod()
              if (isCurrent(object)["MySet"]) return(object)
              ## Create an updated instance.
              if (!isVersioned(object))
                  ## Radical surgery -- create a new, up-to-date instance
                  new("MySet",
                      assayData = updateObject(assayData(object),
                        ...., verbose=verbose),
                      phenoData = updateObject(phenoData(object),
                        ..., verbose=verbose),
                      experimentData = updateObject(experimentData(object),
                        ..., verbose=verbose),
                      annotation = updateObject(annotation(object),
                        ..., verbose=verbose))
              else {
                  ## Make minor changes, and update version by consulting class definition
                  classVersion(object)["MySet"] <-
                      classVersion("MySet")["MySet"]
                  object
              }
          })
@ 
The code after \Rfunction{if(!isVersioned)} illustrates one way of
performing `radical surgery, creating a new up-to-date instance by
updating all slots. The \Rfunction{else} clause represents more modest
changes, using methods to update stale
information. \Rfunction{updateObject} then returns a new, enhanced
object:  
<<updateObject>>=
classVersion(updateObject(obj))
@ 

As in the example, versioning helps in choosing which modifications to
perform -- minor changes for a slightly out-of-date object, radical
surgery for something more ancient. Version information might also be
used in methods, where changing class representation might facilitate
more efficient routines.

\subsection{\Rclass{Versioned} versus \Rclass{VersionedBiobase}}

The information on \R{} and \Biobase{} versions is present in \eSet{}
derived classes because \eSet{} contains \Rclass{VersionedBiobase}. On
the other hand, \Rclass{AnnotatedDataFrame} contains \Rclass{Versioned}, and
has only information about its own class version.
<<classVersion-AnnotatedDataFrame>>=
classVersion(new("AnnotatedDataFrame"))
@ 
The rationale for this is that \Rclass{AnnotatedDataFrame} is and will likely
remain relatively simple, and details about \R{} and \Biobase{} are
probably irrelevant to its use. On the other hand, some aspects of
\eSet{} and the algorithms that operate on them are more cutting edge
and subject to changes in \R{} or \Biobase{}. Knowing the version of
\R{} and \Biobase{} used to create an instance might provide valuable
debugging information.

\subsection{Adding \Rclass{Versioned} information to your own classes}

The key to versioning your own classes is to define your class to
\Rfunction{contain} \Rclass{Versioned} or \Rclass{VersionedBiobase},
and to add the version information in the prototype. For instance, to
add a class-specific version stamp to \Rclass{SwirlSet} we would
modify the class defintion to
<<SwirlSet-version>>=
setClass("SwirlSet", contains = "eSet",
         prototype = prototype(
           new("VersionedBiobase",
               versions=c(classVersion("eSet"), SwirlSet="1.0.0"))))
classVersion(new("SwirlSet"))
@ 
See additional examples in the \Rclass{Versioned} help page. 

It is also possible to add arbitrary information to particular
instances.
<<arbitraryClassVersions>>=
obj <- new("SwirlSet")
classVersion(obj)["MyID"] <- "0.0.1"
classVersion(obj)
classVersion(updateObject(obj))
@ 

There is additional documentation about these classes and methods in
\Biobase{}.

\section{Summary}

This document summarizes \Biobase, outlining strategies that
developers using \Biobase{} may find useful. The main points are to
introduce the \eSet{} class hierarchy, to illustrate how developers
can effectively extend this class, and to introduce class versions as
a way of tracking and easily updating objects. It is anticipated that
\eSet{}-derived classes will play an increasingly important role in
\Biobase{} development.

\section{Session Information}

The version number of \R{} and packages loaded for generating the vignette were:
<<echo=FALSE,results=tex>>=
toLatex(sessionInfo())
@

\end{document}