Full text of DOI:10.1002/cyto.10041

© 2001 Wiley-Liss, Inc.
DOI 10.1002/cyto.10041
Cytometry 47:46–49 (2002)
The Bioinformatics of Microarray Gene
Expression Profiling
*
John N. Weinstein , Uwe Scherf, Jae K. Lee, Satoshi Nishizuka, Fuad Gwadry, Ajay, Kim Bussey,
S. Kim, Lawrence H. Smith, Lorraine Tanabe, Samuel Richman, Jessie Alexander,
Hosein Kouros-Mehr, Alika Maunakea, and William C. Reinhold
Genomics and Bioinformatics Group, Laboratory of Molecular Pharmacology, National Cancer Institute, Bethesda, Maryland
Key terms: gene expression profiling; microarray; biochip; cDNA; oligonucleotide; clustering; clustered image
map Cytometry 47:46–49, 2002. © 2001 Wiley-Liss, Inc.
Gene expression profiling will revolutionize biology.
That much is universally agreed. But it’s harder than it
looks. In part, the reasons can be technical—substandard
arrays, low signal:noise ratios for rare transcripts, variable
backgrounds, cross-hybridizations, the difficulty of pro-
cessing clinical materials, and so forth. But more often the
reasons relate to analysis and interpretation of the data.
Inevitably, more time and energy are spent after the
experiments are finished than before.
have; microarray data analysis is more specialized. What
might be termed developmental bioinformatics involves
the generation of new algorithms (principally by statisti-
cians or those with expertise in machine learning) and the
creation of new software (principally on the basis of
expertise in computer science). Experience shows that
the best analytical developments arise from close atten-
tion to needs arising from actual experimental data sets
and biologic questions
Task #2: To convert images in pixel form to raw
expression levels. Whether one is reading radioisotopi-
cally tagged cDNA in a phosphoimager or measuring flu-
orescent cDNA with a confocal scanner or CCD camera, it
is necessary to develop effective image processing algo-
rithms (See 2, 3). The specifics depend on the type of
array and detection system used and the quality of the
images. As the technologies improve, uncertainties due to
such factors as inhomogeneity in the spots, irregular back-
ground, scanner artifacts, photobleaching, and lack of
spatial registration between channels are diminishing.
Task #3: To examine the array images for quality
control. This important step is facilitated by software
packages that permit surveys of the array image at various
levels of resolution and permit individual spots to be
examined and compared visually.
Task #4: To preprocess the expression-level data (i.e.,
filter, normalize, and/or standardize it). Generally, the
data must be filtered to eliminate flawed spots and genes
with insufficient patterns or differences among samples.
In the former case, it may be necessary, depending on the
nature of the intended analysis, to use statistical or ma-
chine learning techniques to impute values for the missing
data. The next step is normalization, which usually has
been done in the case of two-color studies by tuning a
calibration factor, either on the basis of total gene expres-
sion in the sample or on the basis of a housekeeping gene
We can identify a number of necessary tasks in the
analysis of gene expression data, as summarized in Table
1. In the following capsule descriptions, we will focus for
concreteness on the two-color fluorescence technologies
(1), but analogous steps are pertinent to one-color fluo-
rescence and radioactive detection methods as well. With
apologies to the many scientists who have been innova-
tive in this field, we intend, in this short summary, to
indicate requirements and options rather than to give a
comprehensive review or to apportion credit for the var-
ious contributions. The examples will focus primarily on
studies from our laboratory.
Task #1: To establish the computer hardware, soft-
ware, and personnel infrastructure for handling and
analyzing gigabyte or terabyte databases. There must be
somewhere to put the data, and there must be fluent
systems for pulling information into the stream of analysis.
As data have outgrown Excel (Microsoft, Redmond, WA)
spreadsheets, the most common, but by no means only,
answers have been database packages like Sybase (Sybase
Inc., Emeryville, CA) or Oracle (Oracle Corporation, Red-
wood Shores, CA). Sometimes, however, flat file formats
suffice. For many of the highly multivariate analyses, to be
discussed later, hardware speed and memory become
significant issues. Most important, however, is the hu-
man infrastructure. Applied bioinformatics, broadly con-
strued, is practiced by the biologist who is fluent in the
use of public and proprietary database resources or who
will perform data analyses—preferably under the supervi-
sion of a statistically trained individual. Fluency with da-
tabase resources is something that every biologist should
*Correspondence to: John N. Weinstein, National Institutes of Health,
Bldg 37, Rm 4E-28, 9000 Rockville Pike, Bethesda, MD 20892
E-mail: weinstein@dtpax2.ncifcrf.gov

BIOINFORMATICS OF GENE EXPRESSION
47
Table 1
set. If the samples to be analyzed do not differ from each
other dramatically, the former basis for calibration is prob-
ably preferable. However, scatter plots of the data with
the logarithm of green fluorescence on one axis and that
of red fluorescence on the other often show curvature
such that a unitary calibration factor will not be the best
approach. A number of methods for dealing with this
problem have been developed. Ours, included in a com-
puter program called PreProc (L.H. Smith, et al., manu-
script in preparation), uses Gaussian-windowed moving
averages to fit the curvature (See 4).
Tasks in the Analysis of Microarray-Based
Gene Expression Data
1. Establish the necessary hardware/software/personnel
infrastructure.
2. Convert images in pixels to raw expression levels.
3. Examine the images for quality control.
4. Preprocess the expression level data (filter, normalize,
standardize).
5. Analyze and visualize “high-dimensional” data.
6. Search the literature on genes and gene-gene relationships.
7. Integrate the expression data with other types of
information.
8. Design the study carefully (replicates, controls)—to be
done first.
Additional problems in analysis arise from asymmetries
between the two labeled species in their hybridization
properties. Better results are obtained with reciprocal
FIG. 1. Clustered image map (CIM) relating activity patterns of 118 tested compounds to the expression patterns of 1376 genes in the 60 cell lines of
the National Cancer Institute’s Drug Discovery Program. A red point (high positive Pearson correlation coefficient) indicates that the agent tends to be more
active against cell lines that express more of the gene; a blue point (high negative correlation) indicates the opposite tendency. Genes were cluster-ordered
on the basis of their correlations with drugs (mean-subtracted, average-linkage clustered with correlation metric); drugs were clustered on the basis of their
correlations with genes (mean-subtracted, average-linkage clustered with correlation metric). Insert A shows a magnified view of the region around the
point (white circle) representing the correlation between the dihydropyrimidine dehydrogenase gene (#76) and 5-fluorouracil (#25). Insert B is an
analogous magnified view for the asparagine synthetase gene (#924) and the drug L-asparaginase (#55). These two correlations have led to insights of
potential medical significance (10). Modified from (10).

48
WEINSTEIN ET AL.
averaging, that is, by running replicate arrays in which the
colors are reversed (4). After normalization, a number of
choices must be made before higher level analysis. Should
the data be log transformed to obtain more nearly normal
distribution and heteroskedasticity? Must it then be
thresholded? Should the mean over samples for a given
gene be subtracted from the expression levels? The mean
over genes for a given sample? Should the levels be di-
vided by a measure of the dispersion such as the standard
deviation in one or (by an iterative process) both direc-
tions? Should the continuous values be binned, binarized,
or turned into ranks for analysis? The answers to these
and other such questions often depend on characteristics
of the data or the nature of the question being asked. For
example, if the important information resides in relative,
rather than absolute, gene expression values across a
database, one will likely subtract the mean or median
across samples.
(17), which is freely available (along with databases and
other analytical tools) at http://discover.nci.nih.gov. It
uses a combination of GeneCards from the Weizmann
Institute, PubMed from the National Library of Medicine,
semantic analysis, syntactic analysis, and keywords to find
and organize key sentences from abstracts on genes, gene-
gene relationships, and gene-drug relationships. Med-
Miner can speed up by 5- to 10-fold the rate at which the
voluminous literature on important genes is organized and
interpreted. A version called EDGAR (Extraction of Data
on Genes And Relations) based on deeper semantic anal-
ysis is under development (18).
Task #7: To integrate the expression data with other
types of information. Very often, the gene expression
data are most richly understood and most valuable when
related to other types of information at the protein, DNA,
functional, or pharmacologic level. Figure 1 provides an
example of the pharmacologic connection (10).
Task #5: To analyze and visualize high-dimensional
data. The simplest experimental design is binary—for
example, comparison of cancer with normal cells or ma-
lignant with non-malignant. More complicated is the time
course, for example before and during a treatment. More
demanding still is the large database of samples to be
analyzed for patterns. The latter two types of data are
often presented in the form of what we term clustered
image maps, and others have called heat maps. We intro-
duced clustered image maps (CIMs) for pharmacological,
genomic, and proteomic studies in the mid-1990s (5–7).
Our collaborators later developed a red-black-green color
scheme for CIMs (8, 9). Figure 1 shows a slightly more
complex CIM (10) that relates patterns of gene expression
to patterns of pharmacologic potency in the 60 human
cancer cell lines used in the National Cancer Institute’s
Drug Discovery Program (11, 12). A flexible program for
creating CIMs is available at our web site, http://discover.
nci.nih.gov
Depending on the questions to be asked, high-dimen-
sional data sets may be analyzed by supervised or unsu-
pervised methods. The former include, for example, tech-
niques based on regression, discrimination, or prediction;
the latter on techniques such as clustering (5, 6, 9, 10, 13),
principal components analysis, or multidimensional scal-
ing. There is no right method of analysis. Demands of the
data and the scientific questions asked will condition the
choice.
Task #6: To search the biomedical literature and pub-
lic databases for information on genes or gene-gene
relationships. Most gene expression microarray experi-
ments produce long lists of genes with possible signifi-
cance, and the problem is to distinguish causally interest-
ing relationships from epiphenomenal ones and from
statistical coincidence. For that purpose, outside informa-
tion is generally necessary. Microarray studies are a form
of omic research (14–16), but interpretation of the data
from them generally requires synergy with classical hy-
pothesis-driven studies of one gene, one gene product, or
one process at a time. To facilitate searches of the litera-
ture in this context, we developed the program MedMiner
Task #8: To design the study carefully (in terms of
controls, replicates, internal standards, and design
points). This step should come first, of course. In microar-
ray studies, it is often not feasible to go back afterward and
fill in the gaps in an imperfectly designed or executed
experimental series. Because arrays are expensive, the
tendency is to skimp on replicates and controls, but that
is almost always a mistake. Some of the best and most
often-used databases placed in the public domain to date
suffer from these insufficiencies. Even if it is not practical
to use sufficient replicates for all samples, selected repli-
cates (and replicated genes on each array) pay major
dividends.
This whirlwind summary of the tasks involved in anal-
ysis of microarray gene expression data has by no means
touched on all of the important ingredients of the prob-
lem, let alone presented them in satisfactory detail. More
important than the details of method, however, are com-
mon sense and an appreciation of basic statistical princi-
ples. Artificial intelligence may one day produce software
that can substitute for the human judgment and expertise
currently required for gene expression analysis. But such
software would look nothing like what is now available.
LITERATURE CITED
1. Schena M, Shalon D, Davis RW, Brown PO. Quantitative monitoring
of gene expression patterns with a complementary DNA microarray.
Science 1995;270:467–470.
2. Chen Y, Dougherty ER, Bittner ML. Ratio-based decisions and the
quantitative analysis of cDNA microarray images. J Biomed Optics
1997;2:364–374.
3. Ermolaeva O, Rastogi M, Pruitt KD, Schuler GD, Bittner ML, Chen Y,
Simon R, Meltzer P, Trent JM, Boguski MS. Data management and
analysis for gene expression arrays. Nat Genet 1998;20:19–23.
4. Zhou Y, Gwadry FG, Reinhold WC, Miller L, Smith LH, Scherf U, Liu
E, Kohn KW, Pommier Y, Weinstein. Transcriptional regulation of
mitotic genes by camptothecin-induced DNA damage: Microarray
analysis of dose- and time-dependent effects. Cancer Res Submitted.
5. Weinstein JN, Myers TG, Buolamwini J, Raghavan K, van Osdol W,
Licht J, Viswanadhan VN, Kohn KW, Rubinstein LV, Koutsoukos AD,
Zaharevitz D, Grever MR, Monks A, Scudiero DA Chabner BA Ander-
son NL, Paull KD. Predictive statistics and artificial intelligence in the
U.S. National Cancer Institute’s Drug Discovery Program for Cancer
and AIDS. Stem Cells 1994;12:13–22.
6. Weinstein JN, Myers TG, O’Connor PM, Friend SH, Fornace AJ, Kohn
KW, Fojo T, Bates SE, Rubinstein LV, Anderson NL, Buolamwini JK,

BIOINFORMATICS OF GENE EXPRESSION
49
van Osdol WW, Monks AP, Scudiero DA, Sausville EA, Zaharevitz DW,
Bunow B, Viswanadhan VN, Johnson GS, Wittes RE,Paull KD. An
information-intensive approach to the molecular pharmacology of
cancer. Science 1997;275:343–349.
7. Myers TG, Anderson NL, Waltham M, Li G, Buolamwini JK, Scudiero
DA, Paull KD, Sausville EA, Weinstein JN. A protein expression data-
base for the molecular pharmacology of cancer. Electrophoresis
1997;18:647–653.
8. Eisen MB, Spellman PT, Brown PO, Botstein D. Cluster analysis and
display of genome-wide expression patterns. Proc Natl Acad Sci U S
A 1998;95:14863–14868.
9. Ross DT, Scherf U, Eisen MB, Perou CM, Rees C, Spellman P, Iyer V,
Jeffrey SS, Van de Rijn M, Waltham M, Pergamenschikov A, Lee JC,
Lashkari D, Shalon D, Myers TG, Weinstein JN, Botstein D, Brown PO.
Systematic variation in gene expression patterns in human cancer cell
lines. Nat Genet 2000;24:227–235.
10. Scherf U, Ross DT, Waltham M, Smith LH, Lee JK, Tanabe L, Kohn
KW, Reinhold WC, Myers TG, Andrews DT, Scudiero DA, Eisen MB,
Sausville EA, Pommier Y, Botstein D, Brown PO, Weinstein JN. A gene
expression database for the molecular pharmacology of cancer. Nat
Genet 2000;24:236–244.
11. Boyd MR. Status of the NCI preclinical antitumor drug discovery
screen. In: DeVita VT, Hellman S, Rosenberg SA, editors. Cancer:
principles and practice of oncology update. Philadelphia: Lippincott;
1989. p 1–12.
12. Paull KD, Shoemaker RH, Hodes L, Monks A, Scudiero DA, Rubinstein
L, Plowman J, Boyd MR. Display and analysis of patterns of differential
activity of drugs against human tumor cell lines: Development of
mean graph and COMPARE algorithm. J Natl Cancer Inst 1989;81:
1088–1092.
13. Michaels GS, Carr DB, Askenazi M, Fuhrman S, Wen X, Somogyi R.
Cluster analysis and data visualization of large-scale gene expression
data. Pac Symp Biocomput 1998;:42–53.
14. Weinstein JN. Fishing expeditions. Science 1998;282:628–629.
15. Weinstein JN, Buolamwini JK. Molecular targets in cancer drug dis-
covery: Cell-based profiling. Curr Pharm Design 2000;6:473–483.
16. Weinstein JN. Pharmacogenomics: Teaching old drugs new tricks.
N Engl J Med 2000;343:1408–1409.
17. Tanabe L, Scherf U, Smith LH, Lee JK, Hunter L, Weinstein JN.
MedMiner: an Internet text-mining tool for biomedical information,
with application to gene expression profiling. Biotechniques 1999;
27(6):1210–4, 1216–7.
18. Rindflesch TC, Tanabe L, Weinstein JN, Hunter L. EDGAR: Extraction
of drugs, genes and relations from the biomedical literature. Pac
Symp Biocomput 2000;:517–528.