BIOINFORMATICS APPLICATIONS NOTE
Vol. 23 no. 5 2007, pages 639–640
doi:10.1093/bioinformatics/btl644
Gene expression
Software package for automatic microarray image
analysis (MAIA)
Eugene Novikov and Emmanuel Barillot
Service Bioinformatique, Institut Curie, 26 Rue d’Ulm, 75248 Paris Cedex 05, France
Received on May 11, 2006; revised on December 1, 2006; accepted on December 18, 2006
Advance Access publication January 19, 2007
Associate Editor: Alfonso Valencia
ABSTRACT
Summary: Although various software solutions are currently available for microarray image analysis, one would still expect to develop
algorithms ensuring higher level of intelligence and robustness. We
present a fully functional software package for automatic processing
of the two-color microarray images including spot localization,
quantification and quality control. The developed algorithms aim
at making ratio estimates more resistant to array contamination and
offer automatic tools to evaluate spot quality.
Availability: A demo version of the software can be downloaded
from http://bioinfo.curie.fr/projects/maia. A full version is freely
available to non-commercial users upon request from the authors.
Contact: [email protected]
Two-color (typically, Cy3/green and Cy5/red) comparative
microarray experiment is a key point of cDNA (Hedge et al.,
2000), CGH (comparative genome hybridization, Pinkel et al.,
1998) and, more recently, protein (Eckel-Passow et al., 2005)
microarray technologies. In a conventional two-color experiment, one obtains two images (one for each color) containing
thousands of bright spots, distributed over the array according
to a special pattern. The first step in microarray data analysis
consists of image processing intending to estimate the ratio
of the fluorescence intensities in two color channels at each
spot. This ratio reflects differential gene (cDNA) or protein
expression or change in DNA copy number (CGH) between
two compared samples. Although various academic and
commercial software solutions for microarray image analysis
are currently available (Jain et al., 2002; Molecular Devices,
2005; White et al., 2005; Yang et al., 2002), further efforts will
perhaps never be exhausted because the existing solutions are
not ideal and offer some room for further improvements.
In this note, we present a fully functional software package
(MAIA) for automatic processing of the microarray two-color
images including spot localization, quantification and quality
control (Fig. 1).
The spot localization module (i) identifies the position of each
spot on the array to associate it with the spotted clone;
and (ii) establishes the borders between the neighboring spots
to allow further independent data processing (extracting
*To whom correspondence should be addressed.
quantitative information) for each spot. We deal with the most
widespread, orthogonal, spot localization pattern. In this pattern,
the spots are aligned horizontally and vertically and can be
arranged in blocks containing different numbers of spot rows and
spot columns (Fig. 1). The spot localization algorithm (Novikov
and Barillot, 2006) is fully automatic and robust with respect to
deviations from perfect spot alignment and contamination. As an
input, it requires only the common array design parameters:
number of blocks and number of spots in the x and y directions
of the array. Although fully automatic, there is no guarantee that
it will perform well for any array. Therefore, we offer some
interactive tools to repair grid in case if it is erroneous.
The spot quantification module generates two sets of quantitative parameters for each spot: ratio estimates and spot quality
characteristics. Two algorithms for ratio estimation were
implemented (Novikov and Barillot, 2005a). The first is a
direct arithmetic ratio of the background-corrected fluorescence
intensity estimates in the two color channels. This approach
requires the identification of both the foreground—the measured spot—and the background—typically the level of nonspecific hybridization. The second ratio estimate is the slope of
the linear regression plot of the pixel intensities in one color
channel versus another one. This approach does not require spot
segmentation, however unavoidable presence of some aberrant
or outlier pixels, occurring even in small quantities, can distract
the regression line and strongly bias the regression ratio.
Therefore we have developed a statistical procedure (Novikov
and Barillot, 2005a), which systematically searches and removes
aberrant or outlier pixels (Fig. 1). This procedure guarantees a
higher level of confidence in the ratio estimates obtained using
linear regression approach. Moreover, after aberrant pixels are
removed the segmentation algorithm also ensures more robust
estimates, and there is a greater agreement in the ratio values
obtained by both methods.
The spot quality module provides a value of spot quality
reflecting the level of confidence in the obtained ratio estimate
at each spot. The unique spot quality value is derived from a set
of nine marginal quality parameters characterizing certain
features of the spot (Novikov and Barillot, 2005b). The
contribution of each quality parameter in the overall quality
is automatically evaluated based on the visual classification of
the spots, or using information available from the replicated
spots, located on the same array (Fig. 1) or over a set of
replicated arrays (Novikov and Barillot, 2005b). Therefore
ß The Author 2007. Published by Oxford University Press. All rights reserved. For Permissions, please email: [email protected]
639
E.Novikov and E.Barillot
Fig. 1. A screenshot of the MAIA processing window. An exemplary spot with the pixel regression outliers is shown. Five outlier pixels (in cyan),
which can hardly be flagged out by geometrical criteria, can easily be found using linear regression approach. Three replicated spots are placed as
neighbors in a row on this array. Quality plot demonstrates tendency to decrease the overall quality index of the spot replicates ( y-axis) with the
increase of the ratios variability of the replicates (x-axis). ‘Bad’ spots are the spot with the overall quality index beyond the user-defined threshold.
The overall quality is derived from the nine spot quality characteristics, shown in the table.
the developed procedure allows us not only to quantify
spot quality, but also to identify different types of spot
deficiency occurring in microarray technology. The quality
values can be used either directly to flag out some spots with the
quality lower than the user-defined threshold, or in the followup analysis as a weight controlling the contribution/influence of
the obtained ratio estimates.
Image simulation. We have developed a special software
component for Monte-Carlo simulation of the microarray
images (Novikov and Barillot, 2005a). It takes into account the
statistical noise and different types of contamination like nonspecific hybridization and dust. Since in a simulation experiment the true values of the ratios are precisely known, it allows
us to evaluate, to test and to compare different algorithms of
microarray image analysis objectively.
Software. MAIA runs on 95/98/Me/NT/2000/XP Windows
platforms and needs the Java Runtime Environment. Complete
analysis of one 4 Mb image pair (Cy3/Cy5, 7300 spots; each
spot is 45 pixels) takes 5 s on 3.00 GHz PentiumÕ 4 CPU
with 1 GB of RAM.
The validity of the algorithms has been tested and confirmed
on a large set of experimental images from the different
microarray platforms used within our Institute and on a
representative selection of CGH images obtained from the
UCSF Cancer Center.
ACKNOWLEDGEMENTS
We would like to thank our colleagues from the different
laboratories of the Institute Curie: (F. Radvanyi, UMR 144
640
CNRS/IC: Compartimentation et dynamique cellulaires;
O. Delattre, INSERM 509: Pathologie moléculaire des cancers;
M. Dutreix, UMR 2027 CNRS/IC: Génotoxicologie et cycle
cellulaire) and Prof D. Pinkel (UCSF Comprehensive Cancer
Center), who has provided numerous microarray images
allowing considerable improvement of the algorithms.
Conflict of Interest: none declared.
REFERENCES
Eckel-Passow,J.E. et al. (2005) Experimental design and analysis of
antibody microarrays: applying methods from cDNA arrays. Cancer Res.,
65, 2985–2989.
Hegde,P. et al. (2000) A concise guide to cDNA microarray analysis.
BioTechniques, 29, 548–562.
Jain,A.N. et al. (2002) Fully automated quantification of microarray image data.
Genome Res., 12, 325–332.
Molecular Devices, Corp (2005) GenePix Pro 6.0. http://www.moleculardevices.
com/, User’s Guide and Tutorial.
Novikov,E. and Barillot,E. (2005a) An algorithm for automatic evaluation of the
spot quality in two-color DNA microarray experiments. BMC Bioinformatics,
6, 293.
Novikov,E. and Barillot,E. (2005b) A robust algorithm for ratio estimation in
two-color microarray experiments. J. Bioinform. Comput. Biol., 3, 1411–1428.
Novikov,E. and Barillot,E. (2006) A noise-resistant algorithm for grid finding in
microarray image analysis. Machine Vis. Appl., 17, 337–345.
Pinkel,D. et al. (1998) High resolution analysis of DNA copy number
variation using comparative genomic hybridization to microarrays.
Nat. Genet., 20, 207–211.
White,A.M. et al. (2005) Automated microarray image analysis toolbox for
MATLAB. Bioinformatics, 21, 3578–3579.
Yang,Y.H. et al. (2002) Comparison of methods for image analysis on cDNA
microarray data. J. Comput. Graph. Stat., 11, 108–136.