\documentclass[11pt, letterpaper]{article}
\usepackage[round]{natbib}
\marginparwidth 0pt
\oddsidemargin 0pt
\evensidemargin 0pt
\marginparsep 0pt
\topmargin 0pt
\textwidth 6.75in
\textheight 8.75in
\setlength{\parindent}{0pt}
\setlength{\parskip}{5pt}
\usepackage{moreverb}
\title{MDQC: Mahalanobis Distance Quality Control}
\author{Gabriela V. Cohen Freue and Justin Harrington}
%\VignetteIndexEntry{Introduction to MDQC}
%\VignettePackage{mdqc}
\begin{document}
\maketitle
\tableofcontents
\section{Introduction}
The process of producing microarray data involves multiple steps, some
of which may suffer from technical problems and seriously damage the
quality of the data. Thus, it is essential to identify those arrays
with low quality. Our Mahalanobis Distance Quality Control (MDQC)
is a multivariate quality assessment method for microarrays that is based on the similarity of quality measures across arrays, i.e., on the idea of outlier detection. Intuitively, the ``distance'' of an array's quality attributes measures the similarity of the quality of that array against the quality of the other arrays. Then, arrays with unusually high distances can be flagged as potentially low-quality.
%method uses a multivariate approach to evaluate the quality of an array.
This method computes a distance measure, the Mahalanobis
Distance, to summarize the quality measures contained in a quality
control (QC) report for each array. The use of this distance allows us
to take the correlation structure of the quality measures into
account. In addition, by using robust estimators to identify the
typical quality measures of good-quality arrays, the evaluation is not
affected by the measures of outlying arrays.
We show that computing these distances on subsets of the quality
measures contained in the QC report may increase the method's ability
to detect unusual arrays and helps to identify possible reasons of the
quality problems. Thus, while MDQC can be based on all the quality
measures simultaneously (using \verb+method="nogroups"+ in \verb+mdqc+ function), it is usually
recommended to compute the MDs on subsets of them (using
\verb+method="apriori"+, \verb+"cluster"+, or \verb+"loading"+), or on a transformed space
with a lower dimension (using \verb+method="global"+). In the \verb+"apriori"+
approach the user forms groups of quality measures on the basis of an
a priori interpretation of them and according to the quality aspect
they represent. The \verb+"cluster"+ and the \verb+"loading"+ methods are two
data-driven methods to form the groups. The former groups the quality
measures using clustering analysis, and the latter uses the loadings
of a robust principal component analysis (PCA) to identify the quality measures
that contain similar information and group them. It is important to
note that the \verb+"apriori"+, the \verb+"cluster"+, and the \verb+"loading"+ methods
create groups of the original quality measures of the report and
compute one MD within each group. Finally, the "global" method
computes a single MD based on the reduced space of the first k
principal components from a robust PCA. The number $k$ of PCs can be
chosen using a scree plot.
A robust PCA and the associated biplot and scree plot can be obtained using the prcomp.robust function in this package.
More details on each method are given in \cite{cohen}.
The resulting MDs can be used to flag poor-quality arrays as their MDs
will be large relative to those of undamaged arrays, i.e., they will be
far from the center of the normal arrays. Under usual distributional assumptions, the squared MDs have an approximate Chi-Squared distribution with p degrees of freedom. Thus, using the Chi-Squared distribution we can set a cutoff point to decide if the array is likely to be defective.
Before illustrating the performance of MDQC, we end with some remarks about MDQC. First, it is important to note
that although we use MDQC in the context of quality assessment of microarrays, most of the methodologies implemented
by MDQC have been widely used in Statistics to detect outliers. Thus, although we illustrate our method using two
data sets of Affymetrix GeneChips and their QC reports, all the ideas can be applied to other platforms and/or QC reports
as well as for outlier detection outside microarray data. Second, MDQC can only be used if the number of observations (n)
is greater than or equal to the number of quality measures in the group (p). Thus, if p>n, the user needs to divide the quality measures into
groups (see the apriori, clustering or loading PCA grouping methods in the paper \cite{cohen}) so that each group contains less
variables than observations. Otherwise, other PCA methods have to be applied to reduce the dimensionality of the data set (see \cite{huber}).
Finally, as in any robust estimation method, caution should be taken when MDQC is applied to data sets with small number of arrays. In
these cases, the robust estimators of location and covariance matrix may not be properly scaled damaging the performance of MDQC.
\section{Acute Lymphoblastic Leukemia Study}
To evaluate the performance of MDQC, we use part of an acute lymphoblastic leukemia study described in
\cite{ross}, containing 20 Affymetrix HG-U133Bmicroarrays (see the allQC help file for more details on this data
set). However, MDQC can be used on any QC report for other types of microarrays. \cite{bols} and \cite{bret} examined the quality of these arrays using histograms of probe-level data, MA-plots and probe-level
model methods described in Bioconductor's affyPLM package.
According to their quality assessment, array 2 has a strong spatial artifact on the chip and array 14 presents other evidence of poor quality.
<<>>=
library(mdqc)
data(allQC)
dim(allQC)
allQC[1:2,]
@
As it was previously mentioned, when the number of quality measures is smaller
than the number of observations, one can perform a
\emph{multivariate} analysis using MDQC based on \emph{all} the
quality measures in the report.
<<>>=
mdout <- mdqc(allQC, method="nogroups")
plot(mdout)
print(mdout)
summary(mdout)
@
\begin{figure}[!bp]
\centering
<<fig=TRUE,echo=FALSE>>=
plot(mdout)
@
\caption{Results of MDQC based on all measures of the QC report. The
MDs (y-axis) are computed using the robust $S$-estimator for each
array (x-axis). The solid, dashed and dotted lines correspond to the
square root of the 90th, 95th and 99th percentile of the Chi-Squared
distribution, respectively. Outlying arrays are identified using
solid points.}
\label{noGroups}
\end{figure}
Figure \ref{noGroups} shows that using this MDQC approach array 14 is
flagged as having potential quality problems and array 2 appears only
as a borderline case. In other words, collapsing \emph{all} the quality measures
into a single MD downweights array 2's quality problems and masks
other outlying observations in the QC report, such as those of arrays
1, 7 or 8. Thus, we study the MDs on groups with a reduced number of
variables. These alternative methods reduce the possibility of
masking outliers and may give information about the potential source
of the quality problem.
Based on the interpretation of the quality measures in the QC report,
the user may want to group these measures into the following three
groups:
\begin{enumerate}
\item Scale Factor, \% Present, Avg BG, Min BG, Max BG
\item BioB, BioC, BioD, CreX
\item AFFX-HSAC07/X00351.3'/5', GapDH
\end{enumerate}
This MDQC approach, referred as the \textbf{a priori grouping method},
examines one MD for each array and each group. It is important to recall that
each group can not contain more quality measures than arrays.
<<>>=
mdout <- mdqc(allQC, method="apriori",
groups=list(1:5, 6:9, 10:11))
plot(mdout)
@
\begin{figure}[!tp]%figure2
\centering
<<fig=TRUE,echo=FALSE>>=
plot(mdout)
@
\caption{Results of MDQC using the a priori grouping method. The MDs
(y-axis) within each group are computed using the robust
$S$-estimator for each array (x-axis). The scale of the y-axis
varies from one case to another. The solid, dashed and dotted lines
correspond to the square root of the 90th, 95th and 99th percentile
of the Chi-Squared distribution, respectively. Outlying arrays are
identified using solid points.}
\label{aPriori}
\end{figure}
In Group 1, arrays 2 and 14 are \emph{both} flagged as potentially
defective, and array 17 as a borderline case. In Group 2, array 1 has
a MD exceeding the 99\% cutoff and arrays 7 and 14 have MDs exceeding
the 95\% cutoff line. Finally, array 8 is the only one flagged in
Group 3. Thus, the MDs based on groups of lower dimension flag both
arrays 2 and 14, which is consistent with the results in \cite{bols}
and \cite{bret}. In addition, arrays 1 and 8 are flagged as
potentially low quality and arrays 7 and 17 as borderline
quality. Moreover, based on the interpretability of the groups, the
problems in array 2 are most likely due to defects in the chip as this
array is only identified in Group 1. Similarly, since arrays 1 and 14
are flagged in Group 2, their low quality is most likely due to low
quality of the sample. Note that although array 14 is also flagged in
Group 1, this can be still due to quality problems in the
sample. Finally, array 8 is flagged only in Group 3, suggesting
potential problems in the RNA quality.
While in the apriori grouping method, the groups are created by the
user, the \textbf{clustering grouping method} and the \textbf{loading
PCA grouping method} are two data-driven methods used to create the
groups of quality measures used to compute the MDs. The
\textbf{clustering grouping method} uses the Partitioning Around
Medoids (pam) clustering algorithm based on the distance constructed
from the correlation matrix (i.e., the distance between two variables
x and y is given by d(x,y)=(1-cor(x,y))/2) to group the quality
measures of the QC report. As the data set may contain outlying
observation, using the distance based on the robust estimator of the
correlation matrix as a dissimilarity measure makes it more robust.
We use the S-estimator (\cite{lop}) with 25\% breakdown point to
estimate the correlation matrix of the quality measures. This
estimator has demonstrated a good performance in studies with small
number of arrays, however, other robust estimators can be used (e.g.,
MCD). Although in this method the user can choose the number of clusters (k),
the number of returned quality measures in each group (cluster) can not exceed the
number of arrays in the study. If so, the user needs to increase the number of clusters.
<<>>=
mdout <- mdqc(allQC, method="cluster", k=3)
plot(mdout)
print(mdout)
@
\begin{figure}[!tp]
\centering
<<fig=TRUE,echo=FALSE>>=
plot(mdout)
@
\caption{Results of MDQC using the clustering and the loading grouping methods.}
\label{cluster}
\end{figure}
The \textbf{loading PCA grouping method} uses a robust PCA to group the quality measures
in the report according to their contribution to the first $k$ principal
components. As each loading vector derived from a robust PCA shows
the contribution of each quality variable to the first $k$ PCs, these
vectors can be used to group the quality variables based on the
similarity of their contribution to the model. We can apply a
clustering method, such as pam-algorithm or a hierarchical clustering
method, to group the $p$ quality variables based on their contribution
to the first $k$ principal components. As in previous method,
the number of quality measures in each group (cluster) can not exceed the number of arrays in the study.
Thus, if this is the case, the user needs to increase the number of clusters (k) used in the loading space.
For more details on the robust PCA, see the prcomp.robust function in this package and \cite{cohen}.
<<>>=
mdout <- mdqc(allQC, method="loading", k=3, pc=4)
plot(mdout)
print(mdout)
@
It is interesting to see that this approach leads to the same groups as the clustering grouping method, as given in Figure \ref{cluster}.
Finally, we examine the performance of MDQC using the \textbf{global
PCA method} to reduce the dimensionality of the data. MDQC performs a
robust PCA which requires more observations (arrays) than variables (quality measures).
If this is not the case, the user needs to apply other PCA methods (see for example \cite{huber})
or group the variables using previously described methods. Using the
scree plot, we retain $k=4$ principal components in this
analysis. Figure \ref{global} shows the results of the MDQC when a
single MD is calculated based on the first 4 principal components
derived from a robust PCA based on robustly standardized data (see
\cite{cohen} for more details). We note that this approach still
flags arrays 2, 8 and 14 as having potential quality
problems. However, the first two appear only as borderline cases. In
addition, arrays 1, 7 and 17 are still masked using this method.
<<>>=
mdout <- mdqc(allQC, method="global", pc=4)
plot(mdout)
@
\begin{figure}[!bp]%figure3
\centering
<<fig=TRUE,echo=FALSE>>=
plot(mdout)
@
\caption{Results of MDQC using the global PCA method. The MDs (y-axis)
are computed on the first four principal components for each array
(x-axis). The solid, dashed and dotted lines indicate the square
root of the 90th, 95th and 99th percentile of the Chi-Squared
distribution, respectively. Outlying arrays are identified using
solid points.}
\label{global}
\end{figure}
\section{Conclusion}
The previous example shows that all five approaches of MDQC (i.e., all variables, a priori,
clustering, loading and global PCA) identify the problematic arrays 2
and 14. However, the a priori grouping method outstands the problem of
array 2, unmasks other potentially low quality arrays and provides
possible explanations of the quality problems.
In summary, MDQC has a clear statistical foundation, it performs a robust multivariate analysis of the quality measures provided in the QC report while taking into account their correlation structure, it is easy to apply, and it is computationally lightweight. These properties make MDQC a useful diagnostic technique suitable for large data sets.
\begin{thebibliography}{}
\bibitem[Bolstad {\it et~al}., 2005]{bols} Bolstad, B. M.; Collin, F.;
Brettschneider, J.; Simpson, K.; Cope, L.; Irizarry R. A.; and Speed
T. P. (2005) Quality assessment of Affymetrix GeneChip data. In
Gentleman, R.; Carey, C. J.; Huber, W.; Irizarry, R. A.; and Dudoit,
S. \emph{Bioinformatics and Computational Biology Solutions Using R
and Bioconductor}. New York: Springer.
\bibitem[Brettschneider {\it et~al}., 2007]{bret} Brettschneider, J.;
Collin, F.; Bolstad, B. M.; and Speed, T. P. (2007) Quality
assessment for short oligonucleotide arrays. Forthcoming in
\emph{Technometrics} (with Discussion).
\bibitem[Cohen Freue {\it et~al}., 2007]{cohen} Cohen Freue, G. V. and
Hollander, Z. and Shen, E. and Zamar, R. H. and Balshaw, R. and
Scherer, A. and McManus, B. and Keown, P. and McMaster, W. R. and
Ng, R. T. (2007) MDQC: A New Quality Assessment Method for
Microarrays Based on Quality Control Reports.
\emph{Bioinformatics} \textbf{23} 3162 - 3169.
\bibitem[Huber {\it et~al}., 2002]{huber} Huber, M.; Rousseeuw, P.; and Verboven, S. (2002) A fast
method for robust principal components with applications to
chemometrics. \emph{Chemometrics and Intelligent Laboratory
Systems}, \textbf{60}, 101-111.
\bibitem[Lopuha\"{a}, 1989]{lop} Lopuha\"{a}, H. P. (1989) On the
Relation between S-Estimators and M-Estimators of Multivariate
Location and Covariance. \emph{Ann. Statist.}, \textbf{17},
1662-1683.
\bibitem[Ross {\it et~al}., 2003]{ross} Ross, M. E.; Zhou, X.; Song,
G.; Shurtleff, S. A.; Girtman, K.; Williams, W. K.; Liu, H.;
Mahfouz, R.; Raimondi, S. C.; Lenny, N.; Patel, A.; and Downing,
J. R. (2003) Classification of pediatric acute lymphoblastic
leukemia by gene expression profiling. \emph{Blood} \textbf{102},
2951-9.
\end{thebibliography}
\end{document}