%\VignetteIndexEntry{Analysis of high-throughput microscopy-based screens with imageHTS}
%\VignetteKeywords{high-throughput, microscopy-based screens, image processing, high-content screening}
%\VignettePackage{imageHTS}

\documentclass[10pt,a4paper]{article}
\RequirePackage{amsfonts,amsmath,amstext,amssymb,amscd}
\usepackage{graphicx}
\usepackage{verbatim}
\usepackage{hyperref}
\usepackage{color}
\definecolor{darkblue}{rgb}{0.2,0.0,0.4}

\topmargin -1.5cm
\oddsidemargin -0cm   % read Lamport p.163
\evensidemargin -0cm  % same as oddsidemargin but for left-hand pages
\textwidth 17cm
\textheight 24.5cm

\newcommand{\lib}[1]{{\mbox{\normalfont\textsf{#1}}}}
\newcommand{\file}[1]{{\mbox{\normalfont\textsf{'#1'}}}}
\newcommand{\R}{{\mbox{\normalfont\textsf{R}}}}
\newcommand{\Rfunction}[1]{{\mbox{\normalfont\texttt{#1}}}}
\newcommand{\Robject}[1]{{\mbox{\normalfont\texttt{#1}}}}
\newcommand{\Rpackage}[1]{{\mbox{\normalfont\textsf{#1}}}}
\newcommand{\Rclass}[1]{{\mbox{\normalfont\textit{#1}}}}
\newcommand{\code}[1]{{\mbox{\normalfont\texttt{#1}}}}

\newcommand{\email}[1]{\mbox{\href{mailto:#1}{\textcolor{darkblue}{\normalfont{#1}}}}}
\newcommand{\web}[2]{\mbox{\href{#2}{\textcolor{darkblue}{\normalfont{#1}}}}}

\SweaveOpts{keep.source=TRUE,eps=FALSE,include=FALSE}

\begin{document}

\title{Analysis of high-throughput microscopy-based screens with imageHTS}
\author{Gregoire Pau, Xian Zhang, Michael Boutros, Wolfgang Huber\\\email{gregoire.pau@embl.de}}
\maketitle

\tableofcontents

\section{Introduction}

imageHTS is an R package dedicated to the analysis of high-throughput
microscopy-based screens. The package provides a modular and
extensible framework to segment cells, extract quantitative cell features,
predict cell types and browse screen data through web interfaces. Designed
to operate in distributed environments, imageHTS provides a standardized
access to remote screen data, facilitating the dissemination of
high-throughput microscopy-based screens.

In the following, we first show how to use imageHTS to analyse a microscopy-based
RNA interference (RNAi) screen by automated cell segmentation and extraction of
morphological cell features. In a second example, we demonstrate how to access
and analyse data from a remote screen repository.

\section{Analysis of a microscopy-based screen}

The \Robject{kimorph} screen is an RNAi screen where HeLa cells were fixed 48 h after
siRNA transfection and stained for DNA, tubulin and actin. The screen assays 800
siRNAs and is described in \cite{pmid20531400}. In this section, we are analyzing
a 12-well subset of this screen, of reduced image quality (due to package size considerations),
located in the \Robject{inst/submorph} directory of the imageHTS package.

\subsection{Initialization}

In imageHTS, screen data files can be accessed in two locations: in
a local repository, indicated by \Robject{localPath}, or in an optional remote server
designated by \Robject{serverURL}. If a file is not present in the local repository,
e.g. for storage capacity reasons, imageHTS automatically retrieves the corresponding file
from the remote server to the local repository. This dual repository feature is useful
when screen data is stored in a different location from where it is analysed.

After loading the package imageHTS, we initialize an imageHTS object with
\Rfunction{parseImageConf}. The function takes 3
arguments: an imageHTS configuration file and the variables \Robject{localPath} and
\Robject{serverURL}. The imageHTS configuration file, in DCF format, describes
the general screen configuration: where the microscopy images are located and how the
plates and wells are named. We are using the imageHTS configuration file shown
in section~\ref{sec:conf}. A detailed description of the imageHTS configuration
file can be found in the manual pages of \Rfunction{parseImageConf}. We set the
variable \Robject{localPath} to a temporary directory, for storing intermediate
analysis files. The variable \Robject{serverURL} can point either to a directory
or an external URL. In the following example, \Robject{serverURL} points to the
submorph screen data directory of the imageHTS package, which contains the source
images acquired from the microscope. 

<<init, results=hide, echo=FALSE>>=
display = function(...) {invisible()}
@ 

<<library, results=hide>>=
library('imageHTS')
@ 
<<parseImageConf>>=
localPath = tempdir()
serverURL = system.file('submorph', package='imageHTS')
x = parseImageConf('conf/imageconf.txt', localPath=localPath, serverURL=serverURL)
@

The imageHTS object \Robject{x} is now instantiated. The function \Rfunction{configure}
configures the screen by providing the screen description, the plate configuration layout (how sample,
control and empty wells are located in the plates) and the screen log. The
function \Rfunction{annotate} sets up the mapping between reagents and gene targets.
Both functions originate from the package cellHTS2, dedicated to the analysis
of low-content RNAi screens \cite{pmid16869968}. The imageHTS class extends the
cellHTS class and both functions are fully compatible with their cellHTS2 counterparts.
See cellHTS2 documentation for details.

<<configure>>=
x = configure(x, 'conf/description.txt', 'conf/plateconf.txt', 'conf/screenlog.txt')
x = annotate(x, 'conf/annotation.txt')
@

In imageHTS, each well is uniquely referred by an unique ID. Well unique IDs are
generated by the function \Rfunction{getUnames}, which can filter wells
according to their plate, replicate, row, column or content type
(as described in the plate configuration file). The following example enumerates
the wells that are not empty.

<<getUnames>>=
unames = setdiff(getUnames(x), getUnames(x, content='empty'))
unames 
@

12 wells are non-empty in this screen. Metadata (plate, replicate, content, gene
target, annotation) about the wells is retrieved using the function
\Rfunction{getWellFeatures}.

<<getWellFeatures>>=
getWellFeatures(x, unames[1:3])
@

\subsection{Cell segmentation}

Cells present in wells can be segmented using the function \Rfunction{segmentWells}.
\Rfunction{segmentWells} is a high-level function that takes a vector of unique well
IDs and a DCF segmentation parameter file. \Rfunction{segmentWells} uses the
low-level segmentation function indicated by the field \Robject{seg.method} of
the segmentation parameter file to segment individual well images. For each well,
\Rfunction{segmentWells} writes in the local directory: calibrated image data 'cal',
segmentation data 'seg' and several JPEG images. Files can be accessed later on with
the functions \Robject{fileHTS} and \Robject{readHTS}, as shown in the sequel.

If an unique well is given, \Rfunction{segmentWells} returns a list of three
images: a calibrated image, a nucleus mask and a cell mask. The images can be
manipulated with the package EBImage \cite{pmid20338898} and visualized using
the command \Rfunction{display}. The function \Rfunction{highlightSegmentation}
merges the calibrated image, the nucleus and cell masks to produce a composite image
that highlights the segmentation information.

In the following, we segment the third negative control well \Robject{rluc} using the
segmentation parameter file shown in section~\ref{sec:conf}. The field \Robject{seg.method}
of the file indicates the function \Robject{segmentATH} to segment the well. This
function is specifically designed to segment cells stained for DNA and cytoskeletal
proteins but any other segmentation function can be used, e.g. for segmenting yeast
cells in bright field images or segmenting organelles stained with specific markers.

<<segmentWells1>>=
uname = getUnames(x, content='rluc')[3]
print(uname)
y =  segmentWells(x, uname=uname, segmentationPar='conf/segmentationpar.txt')
display(y$cal)
hseg = highlightSegmentation(0.6*y$cal, y$nseg, y$cseg, thick=TRUE)
display(hseg)
@ 
<<segmentWells1h,echo=FALSE>>= 
writeImage(y$cal, 'cal.jpeg', quality=80)
writeImage(hseg, 'hseg.jpeg', quality=80)
@

\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.9\textwidth]{cal.jpeg}
\caption{Calibrated image 'y\$cal' from well '001-02-C03'.}
\end{center}
\end{figure}

\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.9\textwidth]{hseg.jpeg}
\caption{Segmented image 'hseg' from well '001-02-C03'. Cell nucleus is
  highlighted in yellow and cell membrane is indicated in magenta.}
\end{center}
\end{figure}

Segmentation of the full screen is done with the following commands and
takes about 4 minutes with a single processor. Since wells can be 
segmented independently from each other, segmentation of the full screen
can be easily parallelized using many processors. The following example
is not run in this vignette, due to time constraints.

<<segmentWells2,eval=FALSE>>=
unames = setdiff(getUnames(x), getUnames(x, content='empty'))
segmentWells(x, unames, 'conf/segmentationpar.txt')
@ 

In imageHTS, all screen data files can be accessed through the function
\Rfunction{fileHTS}, including configuration files, source images, segmentation data, cell
features and JPEG images. \Rfunction{fileHTS} creates paths pointing
to screen data files, using a standardized naming scheme. The
following example shows, for the well indicated by \Robject{uname},
how to get access to first channel of the source image, calibrated image data,
and the JPEG image of the well.

<<fileHTS>>=
fileHTS(x, type='source', uname=uname, channel=1)
fileHTS(x, type='seg', uname=uname)
fileHTS(x, type='viewfull', uname=uname)        
@ 

\subsection{Quantification of cell features}

Quantification of cell features is done by the high-level 
function \Rfunction{extractFeatures} on a set of wells,
using a feature parameter file. Similar to the function
\Rfunction{segmentWells}, \Rfunction{extractFeatures} uses the function
indicated by the field \Robject{extractfeatures.method} of the feature
parameter file to extract cell features. For each well, \Rfunction{extractFeatures}
writes features in the local directory, in tab-separated format.
In the following example, we extract cell features from the well indicated by
\Rfunction{uname}, using the feature parameter file shown in section~\ref{sec:conf}.

<<extractFeatures>>=
extractFeatures(x, uname, 'conf/featurepar.txt')
@ 

Cell features can be accessed using the function \Rfunction{fileHTS}, as described
above. However, for convenience purposes, the function \Rfunction{readHTS} combines
\Rfunction{fileHTS} and reads the corresponding file, according to the specified format
(here, tab-separated). The following example reads the cell feature matrix of
well '001-02-C03'.

<<readFeatures>>=
y = readHTS(x, type='ftrs', uname=uname, format='tab')
dim(y)
y[1:10, 1:7]
@ 

84 cells are present in the well and each cell is described with 180 features.
Cell features include geometrical features, moment-based features, Haralick moments and 
Zernicke features. Cell features are described in the manual pages of the function
\Rfunction{getFeatures} of the package EBImage. Some features have a direct
interpretation, such as \Robject{c.g.s}, which measures the cell area or
\Robject{c.t.m.int}, which quantifies the cell tubulin intensity. In the
following example, we display the distribution of the latter within 
the cells of the well, and identify the cells that have a tubulin 
intensity higher than 1600.

<<histfig, fig=TRUE, height=5>>=
hist(y$c.t.m.int, 20, xlab='Cell tubulin intensity (a.u.)', main='')
abline(v=1600, col=2)
cellid = which(y$c.t.m.int>1600)
print(cellid)
@
\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.6\textwidth]{imageHTS-introduction-histfig}
\end{center}
\caption{Distribution of cell tubulin intensity \Robject{c.t.m.int} in cells of well '001-02-C03'.}
\label{fig:histfig}
\end{figure}

Five cells have a tubulin content higher than 1600. Since rows of
cell feature matrix are synchronised with cell indexes in segmentation masks, 
cells can be easily traced back by loading the segmentation information,
as shown in the following example.

<<traceCell>>=
cal = readHTS(x, type='cal', uname=uname, format='rda')
seg = readHTS(x, type='seg', uname=uname, format='rda')
cseg = rmObjects(seg$cseg, setdiff(1:nrow(y), cellid))
hightub = highlightSegmentation(0.6*cal, cseg=cseg, thick=TRUE)
display(hightub)
@ 
<<traceCellh, echo=FALSE>>= 
writeImage(hightub, 'hightub.jpeg', quality=80)
@

\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.9\textwidth]{hightub.jpeg}
\caption{Cells of well '001-02-C03' having a tubulin intensity \Robject{c.t.m.int} higher than 1600.}
\end{center}
\end{figure}

\subsection{Prediction of cell classes}

Cell features can be used as covariates to classify cells, using
supervised learning and a set of manually annotated cells. The function 
\Rfunction{readLearnTS} takes as arguments a training set file and the feature
parameter file, previously used in \Rfunction{extractFeatures}. The training
set is a list of labelled cells and the feature parameter file contains the
field \Robject{remove.classification.features}, indicating the features that
should not be used during training/classification (e.g. cell position).
Construction of the training set is done using the annotation web module 
cellPicker as described in the section \ref{sec:cellpicker}.

The function \Rfunction{readLearnTS} uses a Support Vector Machine
with a radial kernel to predict cell labels. Training is done by parameter
grid-search and 5-fold cross-validation, to minimize classification error. The
function creates the file \Robject{data/classifier.rda}, which contains the
trained classifier. The following example trains a cell classifier, but is
not run in the vignette due to time constraints.

<<readLearnTS, eval=FALSE>>=
readLearnTS(x, 'conf/featurepar.txt', 'conf/trainingset.txt')
@ 

After training, prediction of cell labels is done by the function
\Rfunction{predictCellLabels}. The function writes for each well
a vector of predicted cell labels. The following example predicts
the cell labels of the well '01-02-C03', using a classifier
previously trained on a set of 184 cells labelled with 3
cell classes: I (interphase), M (mitotic) and D (debris).

<<predictCellLabels>>=
predictCellLabels(x, uname)
@ 

66 interphase, 9 mitotic and 9 debris cells were predicted in the image.
The following example retrieves and displays the predicted cell labels.

<<drawCellLabels>>=
clab = readHTS(x, type='clabels', uname=uname, format='tab')
labid = split(1:nrow(clab), clab$label)
inter = seg$cseg%in%labid$I
mito = seg$cseg%in%labid$M
debris = seg$cseg%in%labid$D
dc = Image(c(inter+mito, inter, debris+inter), colormode='Color', dim=c(dim(seg$cseg)[1:2], 3))
dc = highlightSegmentation(0.5*dc+0.2*drop(cal), cseg=seg$cseg, thick=TRUE)
display(dc)
@ 
<<drawCellLabelsh, echo=FALSE>>= 
writeImage(dc, 'dc.jpeg', quality=80)
@

\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.9\textwidth]{dc.jpeg}
\caption{Predicted cell labels (grey: interphase, red: mitotic, blue: debris) in well '001-02-C03'.}
\end{center}
\end{figure}

Overall prediction is very good, except for few cells. Classification performance
can be easily improved by enlarging the training set and re-run the training and
predicting steps. The cellPicker web module, described in section~\ref{sec:cellpicker},
has an interactive cell annotation interface which is very useful to refine the
training set.

\subsection{Phenotype summarization}

Cell population features are summarized by \Rfunction{summarizeWells}. The
function computes for each well a phenotypic profile, which summarizes
cell population features. Currently, a phenotypic profile consist of:
cell number \Robject{n}, median cell feature \Robject{med.*} (for each feature)
and cell class ratios.
\Rfunction{summarizeWells} creates the file \Robject{data/profiles.tab} which
contains the phenotypic profiles. The following example computes the phenotypic
profiles of all the wells, but is not run in the vignette due to time constraints.

<<summarizeWells, eval=FALSE>>= 
summarizeWells(x, unames, 'conf/featurepar.txt')
@

In the following example, the phenotypic profiles (previously computed and stored in the imageHTS
package) are loaded with \Rfunction{readHTS} and averaged by well type. Only the following
features are considered: n (cell number), med.c.g.s (median cell size), med.c.t.m.int (median cell tubulin
intensity), M (mitotic cell fraction) and D (debris cell fraction).

<<sum>>=
profiles = readHTS(x, type='file', filename='data/profiles.tab', format='tab')
wfcontent = factor(as.character(getWellFeatures(x, unames)$controlStatus))
table(wfcontent)
zwf = split(1:nrow(profiles), wfcontent)
ft = c('n', 'med.c.g.s', 'med.c.t.m.int', 'M', 'D')
avef = do.call(rbind, lapply(zwf, function(z) colMeans(profiles[z, ft])))
print(avef)
@ 

There are 4 \Robject{rluc} negative controls, 4 \Robject{ubc} positive controls
and 4 \Robject{sample} wells in this screen. The average number of cells in
\Robject{ubc} wells is 23.75, lower than in \Robject{rluc} wells, 90.50. 
Moreover, the average fraction of debris cells in \Robject{ubc} wells, 0.59, 
is higher than in \Robject{rluc} wells, 0.04. A larger number of replicates and
proper statistical testing would be needed to determine whether the observed
changes are statistically significant.

\subsection{Configuration files and complete script}
\label{sec:conf}

Configurations files used in this vignette are reproduced in this section.
Since the files are part of the screen data, they can be read using 
\Rfunction{fileHTS}. In the following example, we display the imageHTS configuration file,
the segmentation parameter file and the feature parameter file.

<<imageconf>>=
f = fileHTS(x, 'file', filename='conf/imageconf.txt')
cat(paste(readLines(f), collapse='\n'), '\n')
@ 

<<segmentationpar>>=
f = fileHTS(x, 'file', filename='conf/segmentationpar.txt')
cat(paste(readLines(f), collapse='\n'), '\n')
@ 

<<featurepar>>=
f = fileHTS(x, 'file', filename='conf/featurepar.txt')
cat(paste(readLines(f), collapse='\n'), '\n')
@ 

The following example is the complete script used to automatically segment cells, quantify 
cell features, predict cell labels and summarize phenotypes of the whole screen.
The example is not run in this vignette, due to time constraints.

<<script, eval=FALSE>>=
library('imageHTS')
localPath = tempdir()
serverURL = system.file('submorph', package='imageHTS')
x = parseImageConf('conf/imageconf.txt', localPath=localPath, serverURL=serverURL)
x = configure(x, 'conf/description.txt', 'conf/plateconf.txt', 'conf/screenlog.txt')
x = annotate(x, 'conf/annotation.txt')
unames = setdiff(getUnames(x), getUnames(x, content='empty'))
segmentWells(x, unames, 'conf/segmentationpar.txt')
extractFeatures(x, unames, 'conf/featurepar.txt')
readLearnTS(x, 'conf/featurepar.txt', 'conf/trainingset.txt')
predictCellLabels(x, unames)
summarizeWells(x, unames, 'conf/featurepar.txt')
@

\section{Getting access to remote screen data}
\label{sec:kimorph}

The dual repository architecture of imageHTS allows an easy access to remote screen data.
In the following, we are analysing the full \Robject{kimorph} RNAi screen, targeting
about 800 protein coding genes in HeLa cells. Screen details are available in \cite{pmid20531400}.
The screen has been previously analysed by imageHTS and screen data is located at
\url{http://www.ebi.ac.uk/~gpau/imageHTS/screens/kimorph}. The interactive webQuery browsing
interface is available at \url{http://www.ebi.ac.uk/~gpau/imageHTS/screens/kimorph/webquery}.

\subsection{Initialization}

We first initialize an imageHTS object by setting the variable \Robject{serverURL}
to the screen data URL and the local repository \Robject{localPath} to an empty
local directory. We next configure and annotate the imageHTS objects using the
screen configuration files. The files, absent in the local screen directory, are
automatically downloaded from the remote server.

<<init, results=hide>>=
localPath = file.path(tempdir(), 'kimorph')
serverURL = 'http://www.ebi.ac.uk/~gpau/imageHTS/screens/kimorph'
x = parseImageConf('conf/imageconf.txt', localPath=localPath, serverURL=serverURL)
x = configure(x, 'conf/description.txt', 'conf/plateconf.txt', 'conf/screenlog.txt')
x = annotate(x, 'conf/annotation.txt')
@

\subsection{Inspecting data}
\label{sec:cellpicker}

We enumerate the non-empty wells with \Rfunction{getUnames} and
retrieve metadata about them using \Rfunction{getWellFeatures}. The
\Robject{controlStatus} field contains the well type. We then load
the well phenotypic profiles using \Rfunction{readHTS} in the variable
\Robject{xd}.

<<get>>=
us = setdiff(getUnames(x), getUnames(x, content='empty'))
wfcontent = getWellFeatures(x, us)$controlStatus
table(wfcontent)
xd = readHTS(x, 'file', filename='data/profiles.tab', format='tab')
xd = xd[match(us, xd$uname),]
@ 

There are 1750 non-empty wells in this screen, including 1558 sample experiments and
8 controls, each replicated 24 times. In the following example, we show how the
median cell size \Robject{med.c.g.ss} and median cell eccentricity \Robject{med.c.g.ec}
vary within well types.

<<boxplots, fig=TRUE, width=14, height=7>>=
colors = c('#ffffff', NA, '#aaffff', '#ffaaff', '#ff44aa',  '#aaaaff',  '#aaffaa', 
  '#ff7777', '#aaaaaa', '#ffff77')
par(mfrow=c(1,2))
boxplot(xd$med.c.g.ss~wfcontent, las=2, col=colors, main='Median cell size (a.u.)')
boxplot(xd$med.c.g.ec~wfcontent, las=2, col=colors, main='Median cell eccentricity (a.u.)')
@
\begin{figure}[!h]
\begin{center}
\includegraphics{imageHTS-introduction-boxplots}
\end{center}
\caption{Distribution of median cell size and median cell eccentricity among well types.}
\label{fig:boxplots}
\end{figure}

The boxplots show that the \Robject{ubc} control phenotype is characterized
by small and round cells, the \Robject{clspn} control phenotype is characterized
by large cells and the \Robject{trappc3} control phenotype is characterized
by elongated cells. 

To have a screen-wide overview of the well phenotypes, we draw in the following example
a map of the phenotypic profiles using linear discriminant analysis (LDA),
computed on the on the controls \Robject{rluc}, \Robject{ubc} and \Robject{trappc3}.

<<lda1, fig=TRUE>>=
library(MASS)
z = wfcontent %in% c('rluc', 'ubc', 'trappc3')
ft = 14:50
ld = lda(xd[z, ft], as.character(wfcontent[z]))
py = predict(ld, xd[, ft])
plot(py$x[,1:2])
@ 

%\begin{figure}[!h]
%\begin{center}
%\includegraphics[width=0.45\textwidth]{imageHTS-introduction-lda1}
%\end{center}
%\caption{LDA projection of the phenotypic profiles, computed on the
%  control \Robject{rluc}, \Robject{ubc} and \Robject{trappc3} wells.}
%\label{fig:lda}
%\end{figure}

Two wells stand far away from the other ones. Are they novel phenotypes ?
We identify and display them in the following example.

<<standout>>=
unames = us[which(py$x[,1]>500)]
print(unames)
f = fileHTS(x, type='viewunmonted', spot=3, uname=unames[1])
img1 = readImage(f)[1791:2238,1:448,]
display(img1)
f = fileHTS(x, type='viewunmonted', spot=1, uname=unames[2])
img2 = readImage(f)[1:448,1:448,]
display(img2)
@
<<standouth,echo=FALSE>>= 
writeImage(img1, 'img1.jpeg', quality=80)
writeImage(img2, 'img2.jpeg', quality=80)
@

\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.45\textwidth]{img1.jpeg}
\includegraphics[width=0.45\textwidth]{img2.jpeg}
\caption{Well '001-01-A13' and '002-01-I13' showing staining problems.}
\end{center}
\end{figure}

Wells '001-01-A13' and '002-01-I13' have serious staining problems. This is 
an example how a phenotypic map can be used for quality control. The wells
cannot be used in the analysis and can be flagged in the screen log
configuration file. The LDA plot is now redrawn by adjusting plot limits.

<<lda2, fig=TRUE>>=
plot(py$x[,1:2], xlim=c(-35,25), ylim=c(-20,20), cex=0.3)
z = wfcontent!='sample'
points(py$x[z,1:2], col=1, bg=colors[wfcontent[z]], pch=21)
col = rep(1, length(levels(wfcontent)))
col[2] = NA
legend('topleft', legend=levels(wfcontent), col=col, pt.bg=colors[1:length(wfcontent)],
       pch=21, ncol=2, cex=0.8) 
@

\begin{figure}[!h]
\begin{center}
\includegraphics{imageHTS-introduction-lda2}
\end{center}
\caption{LDA projection of the phenotypic profiles, computed on the
  control \Robject{rluc}, \Robject{ubc} and \Robject{trappc3} wells.}
\label{fig:lda}
\end{figure}

Control wells \Robject{ubc}, \Robject{clspn}, \Robject{rluc} and \Robject{trappc3} are
well separated from each other. Control wells \Robject{plk1} seem to display 
similar phenotypes than the negative control \Robject{rluc}: further inspection
will reveal than the siRNA reagent against \Robject{plk1} did not work in this
experiment.

Several sample wells seem to have strong phenotypes, distant from negative
controls. Further data inspection is facilitated by the
webQuery and cellPicker web modules, which allow interactive browsing and cell
selection/annotation using a web browser. In the following example, the
functions \Rfunction{popWebQuery} and \Rfunction{popCellPicker} open
the corresponding modules. See Fig.~\ref{fig:web} for an overview of the webQuery
and cellPicker web modules.

<<popCellPicker, eval=FALSE>>=
popWebQuery(x)
uname = getUnames(x, content='trappc3')[1]
popCellPicker(x, uname)
@ 

\begin{figure}[!h]
\begin{center}
\includegraphics[width=0.8\textwidth]{webquery.jpeg}\\
\includegraphics[width=0.8\textwidth]{cellpicker.jpeg}
\end{center}
\caption{The webQuery (top) and cellPicker (bottom) web modules.}
\label{fig:web}
\end{figure}

\section{Session info}

This document was produced using:

<<sessionInfo, results=tex, print=TRUE>>=
toLatex(sessionInfo())
@ 

\bibliography{imageHTS}
\bibliographystyle{plain}

\end{document}