L E T T E R S
NATURE BIOTECHNOLOGY ADVANCE ONLINE PUBLICATION 1
We describe an algorithm for discovering regulatory networks
of gene modules, GRAM (Genetic Regulatory Modules), that
combines information from genome-wide location and
expression data sets. A gene module is defined as a set of
coexpressed genes to which the same set of transcription
factors binds. Unlike previous approaches1–5 that relied
primarily on functional information from expression data,
the GRAM algorithm explicitly links genes to the factors that
regulate them by incorporating DNA binding data, which
provide direct physical evidence of regulatory interactions.
We use the GRAM algorithm to describe a genome-wide
regulatory network in Saccharomyces cerevisiae using
binding information for 106 transcription factors profiled
in rich medium conditions data from over 500 expression
experiments. We also present a genome-wide location analysis
data set for regulators in yeast cells treated with rapamycin,
and use the GRAM algorithm to provide biological insights into
this regulatory network.
High-throughput biological data sources hold the promise of revolu-
tionizing molecular biology by providing large-scale views of genetic
regulatory networks. Many genome-wide expression data sets are now
readily available, and typical computational analyses have applied
clustering algorithms to expression data to find sets of coexpressed
and potentially coregulated genes1. Recent approaches have used more
sophisticated algorithms; one group of researchers constructed a
probabilistic model that uses expression data to link regulators to reg-
ulated genes2. Their method relies on the assumption that the expres-
sion levels of regulated genes will depend on the expression levels of
regulators, which is a limitation in cases in which the expression level
of the regulator does not change appropriately (e.g., cases of post-tran-
scriptional modification). Other approaches have combined expres-
sion data with additional information, such as shared DNA binding
motifs or Munich Information Center for Protein Sequences (MIPS)
categories3–5, but the use of these data sources provides essentially only
functional or indirect evidence of genetic regulatory interactions.
These methods cannot reliably distinguish among genes that have sim-
ilar expression patterns but are under the control of different regula-
tory networks (see Supplementary Note online for further details).
Large-scale, genome-wide location analysis for DNA-binding regu-
lators offers a second means for identifying regulatory relationships6.
Location analysis identifies physical interactions between regulators
and DNA regions, providing strong direct evidence for genetic regula-
tion. Although helpful, the usefulness of binding information is also
limited, as the presence of the regulator at a promoter region indicates
binding but not function. The regulator may act positively, negatively
or not at all. In addition, as with all microarray-based data sources,
location analysis data contain substantial experimental noise. Because
expression and location analysis data provide complementary infor-
mation, our goal was to develop an efficient computational method
for integrating these data sources. We expected that such an algorithm
could assign groups of genes to regulators more accurately than meth-
ods based on either data source alone.
The GRAM algorithm begins by performing an efficient, exhaustive
search over all possible combinations of transcriptional regulators
indicated by the DNA-binding data with a stringent criterion for
determining binding. Once a set of genes to which a common set of
transcriptional regulators binds is found, the algorithm identifies a
subset of these genes with highly correlated expression, which serves as
a ‘seed’ for a gene module. The algorithm then revisits the binding data
and, using a relaxed binding criterion, seeks to add additional genes to
the module that are similarly expressed and to which the same set of
transcriptional regulators binds. Our algorithm allows genes to belong
to more than one module. (See the Methods section for a complete
description of the GRAM algorithm.)
The GRAM algorithm was applied to genome-wide location data for
106 transcription factors and over 500 expression experiments (details
on the data used are available in Supplementary Table 1 online). We
identified 106 gene modules, containing 655 distinct genes and regu-
lated by 68 of the transcription factors. Figure 1 presents a visualiza-
tion of these results as a graph with edges between gene modules and
regulators.
The gene modules abstraction allowed us to label regulator-module
edges in the graph to indicate whether there is significant evidence
(P < 0.05) that regulators may be functioning as activators. Because a
gene module provides a link between a set of regulators and the com-
mon expression pattern of a set of genes to which the regulators bind,
we can use the relationship between a regulator’s expression pattern
1MIT Computer Science and Artificial Intelligence Laboratory, 200 Technology Square, Cambridge, Massachusetts 02139, USA. 2Whitehead Institute for Biomedical
Research, Nine Cambridge Center, Cambridge, Massachusetts 02142, USA. 3MIT Department of Biology, 31 Ames Street, Room 68-132, Cambridge, Massachusetts
02139, USA. 4These authors contributed equally to this work. Correspondence should be addressed to D.G. (gifford@mit.edu).
Published online 12 October 2003; doi:10.1038/nbt890
Computational discovery of gene modules and
regulatory networks
Ziv Bar-Joseph1,4, Georg K Gerber1,4, Tong Ihn Lee2,4, Nicola J Rinaldi2,3, Jane Y Yoo2, François Robert2,
D Benjamin Gordon2, Ernest Fraenkel2, Tommi S Jaakkola1, Richard A Young2,3 & David K Gifford1
©
20
03
N
at
ur
e
Pu
bl
is
hi
ng
G
ro
u
p

ht
tp
://
w
w
w.
n
at
ur
e.
co
m
/n
at
ur
eb
io
te
ch
no
lo
gy

L E T T E R S
2 ADVANCE ONLINE PUBLICATION NATURE BIOTECHNOLOGY
and the common expression pattern of genes in a module to infer
whether a regulator acts as an activator. In contrast, the use of genomic
location data alone allows us only to infer the presence of regulators at
promoters, but not to determine the type of interaction. We searched
for activator relationships by examining regulators with expression
profiles that are positively correlated with the expression profiles of
genes in the corresponding modules. Positive correlation indicates that
higher levels of regulator expression correlate with higher levels of
expression of genes in the module and suggests that the transcription
factor positively regulates the expression of genes in the module. We
determined the statistical significance of the activator relationships by
computing correlation coefficients between all transcriptional regula-
tors studied and all gene modules and taking the 5% positive tail of the
distribution of correlation coefficients. Supplementary Table 2 online
presents the 11 activators identified using the method described above.
Ten of these were previously identified in the literature, suggesting that
this analysis produces biologically meaningful results.
Several findings obtained by analysis of the discovered gene modules
suggest that the algorithm identifies biologically relevant groupings of
genes. First, we found that gene modules generally identify groups of
genes that function in a similar biological pathway as defined by the
MIPS functional categorization7 (see Fig. 1 and Supplementary Table 3
online for details). Second, we found the gene modules to be generally
accurate in assigning regulators to sets of genes whose functions are
consistent with the regulators’ known roles. As an example, Gcr1 is a
well-characterized regulator of glucose metabolism8,9; six of the seven
genes identified in the Gcr1 module are enzymes involved in glycolysis
and gluconeogenesis. Additionally, we found that in most cases in
which a gene module is controlled by one or more regulators, there was
previous evidence suggesting that these regulators interact physically or
functionally (see Supplementary Table 4 online). For example, gene
modules identify Hap2-Hap3-Hap4-Hap5, Hap4-Abf1, Ino2-Ino4,
Hir1-Hir2, Mbp1-Swi6 and Swi4-Swi6 interactions. Taken together,
these results provide evidence that the GRAM algorithm identifies not
only biologically related sets of genes, but also relevant factors that are
interacting to control the genes.
Although genome-wide location data alone are potentially useful for
deriving transcriptional regulatory networks, a key feature of the
GRAM algorithm is its ability to compensate for technical limitations
in the location data through the integration of expression data. To
determine binding events in location data, researchers have previously
used a statistical model and chosen a relatively stringent P-value thresh-
old (0.001) with the intention of reducing false positives at the expense
of false negatives6. The GRAM algorithm presents a useful alternative
to using a single P-value threshold to predict binding events, because
our method allows the P-value cutoff to be relaxed if there is sufficient
supporting evidence from expression data. As an example, consider
Hap4, a well-characterized regulator of genes involved in oxidative
phosphorylation and respiration10. The Hap4 modules contain 28
genes that are involved in respiration and show a high degree of coreg-
ulation over the collected expression data sets (Fig. 2). Six of these genes
(PET9, ATP16, KGD2, QCR6, SDH1 and NDI1) would not have been
identified as Hap4 targets using the stringent 0.001 P-value threshold
(P-values range from 0.0011 to 0.0036). Overall, 627 of 1,560 unique
regulator-gene interactions (40%) in the rich medium network discov-
ered by the GRAM algorithm would not have been detected using only
location data and the stringent P-value cutoff.
To further verify the ability of the GRAM algorithm to lower the rate
of false negatives without substantially increasing the rate of false posi-
tives, we performed gene-specific chromatin-immunoprecipitation
(IP) experiments for the factor Stb1 and 36 genes. The profiled genes
were picked randomly from the full set of yeast genes, with representa-
tives selected from four P-value ranges. In these experiments, we found
that Stb1 bound to three additional genes that had P-values between
Figure 1 Rich medium gene modules
network.Visualization of the transcriptional
regulatory network discovered by the GRAM
algorithm as a graph with edges between gene
modules and regulators shows that there are
many groups of connected gene modules and
regulators involved in similar biological processes.
The network consists of 106 modules containing
655 distinct genes regulated by 68 transcription
factors. In most cases in which a gene module
is controlled by one or more regulators, there
was previous evidence suggesting that these
regulators interact physically or functionally (see
Supplementary Table 3 online for details). The
directed arrows point from transcription factors
to the gene modules that they regulate. Blue
arrows indicate discovered activator regulatory
relationships (see Supplementary Table 2 online
and the text for details). Gene modules are
colored according to the MIPS category to
which a significant number of genes belong
(significance test using the hypergeometric
distribution P < 0.005). Modules containing
many genes with unknown function or an
insignificant number belonging to the same
MIPS category are colored black. When the gene
modules discovered by the GRAM algorithm were
compared to results generated using location data
alone, the GRAM algorithm yielded almost three
times as many modules significantly enriched for
genes in the same MIPS category.
©
20
03
N
at
ur
e
Pu
bl
is
hi
ng
G
ro
u
p

ht
tp
://
w
w
w.
n
at
ur
e.
co
m
/n
at
ur
eb
io
te
ch
no
lo
gy

L E T T E R S
NATURE BIOTECHNOLOGY ADVANCE ONLINE PUBLICATION 3
0.001 and 0.01 in the genomic location experiments and had thus been
excluded under the stringent cutoff. The GRAM algorithm identified
all three as genes to which Stb1 binds without adding any additional
genes that were not detected in the gene-specific chromatin-IP experi-
ments (see Supplementary Table 5 and Supplementary Methods
online for full details).
We also expected that the gene modules derived by the GRAM algo-
rithm would improve on the biological relevance of gene groupings that
could be inferred from location data only. Because genes that participate
in the same biological pathway often have similar expression patterns,
and genes in a module share not only a common set of transcription
factors but also similar expression patterns, we expected that genes in
modules would be more likely to be functionally related than sets of
genes identified by location data alone. Indeed, we found that gene
modules derived using the GRAM algorithm were almost three times
more likely to show enrichment for genes in the same MIPS functional
category than were sets of genes derived solely from location data.
Similarly, we expected that genes in modules derived by the GRAM
algorithm would be more likely to show independent evidence of
coregulation by the regulators assigned to the module than would sets
of genes obtained using location data alone. One line of evidence for
such an improvement would be enrichment for specific DNA
sequence motifs. We identified 34 transcriptional regulators that bind
to genes in at least one module and have well-characterized DNA
binding motifs in the Transcription Factor (TRANSFAC) database11.
For each of these 34 transcriptional regulators, we constructed two
lists of genes, the first using modules to which the regulator binds
(generated by the GRAM algorithm) and the second using location
data alone (stringent P-value cutoff of 0.001). We then computed from
each list the percentage of genes that contained the appropriate known
motif in the upstream region of DNA. We found that in most cases the
percentage of genes containing the correct motif was higher when we
used modules generated using the GRAM algorithm than when we
used sets of genes generated from location data alone (see Fig. 3 and
Supplementary Table 6).
The use of a very large set of genome-wide location and expression
data allowed us to validate the results of the GRAM algorithm compre-
hensively for the gene modules discussed above through literature
searches, independent chromatin-IP experiments, and analysis for
enrichment for genes in the same MIPS category and for known DNA-
binding motifs. The results of this large-scale validation gave us confi-
dence that the GRAM algorithm would be useful in analyzing new data
sources. Because biological insights are often gained by examining
responses to specialized treatments or environmental conditions, we
were interested in exploring the performance of the GRAM algorithm
on a data set that was smaller and more biologically targeted than the
rich medium data. So, we chose to examine a transcriptional regula-
tory subnetwork involved in the response to Tor kinase signaling.
The Tor proteins are highly conserved and function as critical regu-
lators in the response to nutrient stress12–15. Tor kinase signaling can
be inhibited by the addition of the small macrolide rapamycin, which
mimics nutrient starvation and results in a wide range of physiological
responses including cytoskeleton reorganization, decreased transla-
tion initiation, decreased ribosome biogenesis, amino acid permease
regulation and autophagy16–19. Expression analysis indicates that Tor
signaling also controls transcriptional regulation of metabolic path-
ways involving nitrogen metabolism, glycolysis and the tricarboxylic
acid (TCA) cycle15–17.
The rapamycin response presented an ideal opportunity for applying
the GRAM algorithm to the analysis of a novel transcriptional regula-
tory subnetwork. Previous studies suggest a specific set of regulators that
are likely to function in the transcriptional response to rapamycin15,16.
Also, several publicly available genome-wide expression data sets meas-
uring response after rapamycin treatment are available15,16. More
importantly, the fact that there is little information available about the
transcriptional regulatory network involved and how this transcrip-
tional network may contribute to the overall response to rapamycin
treatment presented an opportunity for new biological insights.
We selected 14 transcriptional regulators that seemed likely to func-
tion in the rapamycin response in S. cerevisiae based on evidence from
Figure 2 The GRAM algorithm integrates genome-wide binding and
expression data and improves on either data source alone. (a) Binding data:
the GRAM algorithm can improve the quality of DNA-binding information
because it uses expression data to avoid a strict statistical significance
threshold. Shown is DNA-binding and expression information for the
99 genes bound by the regulator Hap4 with a P value < 0.01 using an
earlier statistical model6. The blue-white column on the left indicates
binding P values, and the horizontal yellow line denotes the strict
significance threshold of 0.001. As can be seen, the P values form a
continuum and a strict threshold is unlikely to produce good results. The
blue horizontal lines on the right indicate the 28 genes that were selected
for modules by the GRAM algorithm. As can be seen, 22 (79%) have a
P value < 0.001, but 6 (21%) have P values above this threshold. The
lower portion of the figure shows together the 28 genes selected by the
GRAM algorithm, and it can be seen that they exhibit coherent expression.
Further, all the selected genes are involved in respiration. Six of these
genes (PET9, ATP16, KGD2, QCR6, SDH1 and NDI1) would not have
been identified as Hap4 targets using the stringent 0.001 P-value
threshold (P values range from 0.0011 to 0.0036). (b) Expression data:
the GRAM algorithm can assign different regulators to genes with similar
expression patterns that cannot be distinguished reliably using expression
clustering methods alone. Hierarchical clustering of expression data was
used to obtain the subtree on the left. On the right, the regulators assigned
to genes by the GRAM algorithm are color coded. As can be seen, many
genes with very similar expression patterns are regulated by different
transcription factors.
a
b
©
20
03
N
at
ur
e
Pu
bl
is
hi
ng
G
ro
u
p

ht
tp
://
w
w
w.
n
at
ur
e.
co
m
/n
at
ur
eb
io
te
ch
no
lo
gy

L E T T E R S
4 ADVANCE ONLINE PUBLICATION NATURE BIOTECHNOLOGY
the literature, and performed genome-wide location analysis experi-
ments (see Methods and Supplementary Table 7 online for full
details). We ran the GRAM algorithm using the location data for the
14 transcription factors in rapamycin and 22 previously published
expression experiments relevant to rapamycin conditions. We discov-
ered 39 gene modules containing 317 unique genes and regulated by
13 transcription factors (see Fig. 4 and Supplementary Table 8 online).
The GRAM algorithm added 192 pairs of gene-regulator interactions
that would not have been identified with a strict P value (0.001) in the
location analysis experiments. Because genome-wide binding experi-
ments for the rapamycin regulatory network have not been performed
before, it was not possible to verify these interactions comprehensively
using literature searches.
As with the rich medium gene modules network, the rapamycin reg-
ulatory network discovered by the GRAM algorithm had many fea-
tures that were consistent with expectations from the literature.
Twenty-three of the gene modules were found to contain a significant
number of genes (P < 0.05) belonging to a single MIPS category. There
were a total of nine categories, all corresponding to biological
responses associated with rapamycin treatment12–14. We also found
that, in general, regulators were assigned to genes that reflect functions
described in previously published results.
In addition to identifying established regulatory interactions, analy-
sis of the rapamycin gene modules suggested several unexpected inter-
actions in which regulators typically assigned to a particular biological
response also appear to bind genes acting in different biological path-
ways. Below we give several examples of such regulatory interactions.
These findings suggest models of transcriptional regulation of the
rapamycin response that can be validated in further, more directed
studies. A first example of an unexpected regulatory interaction
involves the factors Msn2 and Msn4, which are generally regarded as
stress response factors and have been well studied as activators of stress-
related responses18–21. Unexpectedly, there were three gene modules in
which Msn2 and Msn4 bound to a significant number of genes
involved in the mating pheromone response pathway (P < 0.006). A
second example involves the factor Rtg3, which is generally thought to
regulate directly genes of the TCA cycle and indirectly contribute to
nitrogen metabolism22–25 (products of the TCA cycle are shunted to
nitrogen metabolism pathways in low- or poor-nitrogen conditions).
The gene modules network suggests that Rtg3 may directly regulate
genes involved in amino acid metabolism, and more specifically in
nitrogen metabolism.
A third example of an unexpected regulatory interaction involves
Hap2, a part of the Hap2-Hap3-Hap4-Hap5 complex that has been
well characterized as a regulator of genes involved in respiration22,26.
Indeed, in the rich medium gene modules network, members of the
Hap complex are unique among the 106 regulators profiled as the only
regulators controlling modules that are significantly enriched for genes
involved in respiration (P < 0.005). As expected, Hap2 regulates a mod-
ule of respiration genes under rapamycin conditions. Unexpectedly,
Hap2 was also found to regulate two modules containing genes
involved in nitrogen metabolism. There is some genetic evidence for
such cross-pathway regulation, as Hap2 was previously implicated as a
regulator of two nitrogen metabolism genes27,28. Our results indicate
that Hap2 participates in cross-pathway regulation more extensively
than previously reported.
In addition to suggesting that some transcriptional regulators may
control genes involved in biological pathways different from those
Figure 3 Motif enrichment. Genes in modules discovered by the GRAM
algorithm are more likely to show independent evidence of coregulation by
the regulators assigned to the module when compared to sets of genes
obtained using genomic location analysis data alone, as demonstrated by an
enrichment for the presence of known DNA-binding motifs. We identified
34 transcriptional regulators that bind to genes in at least one module and
have well-characterized DNA binding motifs in the TRANSFAC database11.
For each of these 34 transcriptional regulators, we generated a list of genes
in modules bound by the regulator and a second list of genes bound by the
regulator using location analysis data alone (stringent P value cutoff of
0.001). We then computed the percentage of genes from each list that
contained the appropriate known motif in the upstream region of DNA. In
most cases, the percentage of genes containing the correct motif was higher
when we used modules generated by the GRAM algorithm than when we used
sets of genes generated by location analysis data alone. See Supplementary
Table 6 online for a complete list of transcription factors analyzed.
Figure 4 Rapamycin gene modules network. Analysis of the rapamycin
transcriptional regulatory subnetwork revealed a number of novel biological
insights, including evidence that some transcriptional regulators may
control genes involved in biological pathways different from those generally
associated with these regulators. Further, analysis of the network suggested
more complex regulatory interactions in which there is communication
among modules. Such complicated network topologies may be important
for facilitating rapid and flexible responses to changing environmental
conditions. See the text for further details. Thirty-nine modules containing
317 unique genes and regulated by 13 transcription factors were
discovered. Red arrows between transcriptional regulators indicate that the
source transcription factor binds at least one module containing the target
transcription factor. Modules are colored according to the MIPS category to
which a significant number of genes belong (significance test using the
hypergeometric distribution P < 0.05).
©
20
03
N
at
ur
e
Pu
bl
is
hi
ng
G
ro
u
p

ht
tp
://
w
w
w.
n
at
ur
e.
co
m
/n
at
ur
eb
io
te
ch
no
lo
gy

L E T T E R S
NATURE BIOTECHNOLOGY ADVANCE ONLINE PUBLICATION 5
generally associated with these regulators, analysis of the gene modules
network suggests more complex regulatory interactions in which there
is communication among gene modules. Such complicated network
topologies may be important for facilitating rapid and flexible
responses to changing environmental conditions. As an example, we
found that several transcriptional regulators may be involved in a feed-
forward regulatory loop in which the gene encoding a regulator is
bound by another regulator and both regulators bind to a set of com-
mon genes6,29. The regulator Gat1 has been previously identified as a
general activator of nitrogen-responsive genes30. We found that Gat1 is
itself contained in several modules along with genes involved in nitro-
gen metabolism. The transcriptional regulators Dal81, Dal82, Gln3
and Hap2 bind to these gene modules. Interestingly, Gat1 also binds to
several gene modules along with Dal81, Dal82 and Gln3 (see Fig. 4).
Feed-forward mechanisms may be important in regulatory responses
(such as the response to rapamycin) by modulating regulatory sensi-
tivity to sustained rather than transient inputs, providing temporal
control or amplifying the transcriptional response29. These findings
can be validated in further directed experimental studies.
The above analyses indicate that the GRAM algorithm can be useful
for studying transcriptional regulatory networks using genome-wide
location and expression data sources. We have made a Java implemen-
tation of the algorithm publicly available (see Supplementary
Methods online), and believe that as new genome-wide location data
become increasingly available, other researchers will find the algo-
rithm helpful. As demonstrated, the algorithm can integrate sources of
genome-wide location and expression data to help compensate for
technical limitations in the data. Further, the inferred gene modules
networks can give a clearer view of regulation than can either location
or expression data sources alone. We have found that the algorithm is
particularly useful for uncovering how certain regulators may act in
multiple biological pathways. Overall, the GRAM algorithm facilitates
a genome-wide approach to analysis of transcriptional regulatory net-
works that can suggest specific novel regulatory models, which can
then be validated in more directed experimental studies.
METHODS
The GRAM (Genetic Regulatory Modules) algorithm. Below we describe the
operation of the algorithm. Some details are omitted owing to space con-
straints; see the Supplementary Methods online for complete information as
well as a Java implementation of the algorithm.
Let ei denote an expression vector and bi a vector of binding P values for gene
i, where there are ng genes. Let B(i,t) denote the set of all transcription factors
that bind to gene i with a P value less than t, that is, the list of indices j such that
bij < t. Let F ⊆ B(i,t) denote a subset of the transcription factors that bind to i.
Let G(F,t) be the set of all genes i such that for any gene i ∈ G(F,t), F ⊆ B(i,t),
that is, genes to which all the factors in F bind with a given significance thresh-
old. The algorithm begins by going over all genes, and assigning each gene i to
all possible sets G(F,t), where t1 is a high-stringency binding threshold and F
ranges over all subsets of B(i,t).
For every set of transcription factors F, the genes in G(F,t1) serve as candi-
dates for a module regulated by F. For each such set G(F,t1) with a sufficient
number n of genes (e.g., n ≥ 5), the algorithm attempts to find a ‘core’ expression
profile. That is, we are seeking a point c in expression space such that for an
expression similarity threshold sn, the ball centered at c of radius sn contains as
many genes in G(F,t1) as possible. Denote by C(F,t1,c) the ‘core’ set of genes such
that C(F,t1,c) ⊆ G(F,t1) and for each gene i ∈C(F,t1,c), d(ei,c) < sn, where d is the
Euclidian distance between two points. The threshold sn is determined by using
all genes, and randomly sampling subsets of size n to determine the distribution
of expression distances from a subset to all genes. The problem of finding a
point c for a set of expression vectors is nontrivial, and cannot be optimally
solved in a reasonable time given the dimensionality of the expression space
(>500). Thus, we use a theoretically motivated approximation algorithm that
looks for the central point in all triplets of genes in G(F,t1) (see Supplementary
Methods online for more details).
The genes in C(F,t1,c) are used to initialize a module M(F). Conceptually, we
would like to expand this module by relaxing our criteria for binding if a gene’s
expression profile is sufficiently similar to those in the ‘core.’ To do so, the algo-
rithm calculates a combined P value pi for each gene i that belongs to the
expanded set C(F,t2,c) and does not belong to C(F,t1,c), where t2 > t1. The P
value pi is arrived at by computing independent P values for gene i and each
transcription factor in F and then combining the P values using the Fisher
method. A gene i from C(F,t2,c) is then included in M(F) if pi < t1. This module
initialization and expansion is completed for each feasible F, starting with the
sets containing the largest number of factors and proceeding to the smallest. If a
gene is included in a module M(F), it is masked out (not considered) when
forming modules with factor subsets, M(F ′) where F′ ⊆ F. That is, the algorithm
will seek to explain a gene’s expression using the most specific regulatory pat-
terns. The thresholds t1 = 0.001 and t2 = 0.01 were chosen based on experi-
ments6 that suggested very low false positive rates for a significance threshold of
0.001. Further, the rate of false negatives was found to be relatively high for P
values between 0.01 and 0.001, but decreased markedly (to <3%) thereafter.
Strains. Epitope-tagged strains were generated as described6. Briefly, regulators
were tagged at the C terminus by using homologous recombination to insert
multiple copies of the Myc epitope coding sequence into the normal chromoso-
mal loci of these genes. Insertion of the epitope coding sequence was confirmed
by PCR and expression of the epitope-tagged protein was confirmed by western
blotting analysis.
Growth conditions. Strains containing epitope-tagged regulators were grown in
50 ml YPD broth (yeast extract, peptone, dextrose) at 30 °C. Cells were grown to
an OD600 of 0.7–0.8 and rapamycin was then added to a final concentration of
100 nM. Cells were grown for 20 min at 30 °C in the presence of rapamycin.
Genome-wide location analysis. Genome-wide location analysis was done as
previously described6. Briefly, cells containing an epitope-tagged regulator were
fixed with formaldehyde (1% final concentration) and then harvested by cen-
trifugation. Cells were lysed and then sonicated to shear DNA. DNA fragments
representing chromosomal regions crosslinked to a protein of interest were
enriched by immunoprecipitation with an anti-epitope antibody. After reversal
of crosslinking, enriched DNA was purified. The ends of DNA fragments were
then blunted using T4 DNA polymerase and ligated to previously prepared
linkers. The enriched DNA was then amplified and labeled with a fluorescent
dye by ligation-mediated PCR. A sample of control DNA was similarly
processed and labeled with a different fluorophore. Both IP-enriched and con-
trol DNA were then hybridized to a single DNA microarray. For each factor,
three independently grown cell cultures were processed and scanned to gener-
ate binding information as previously described (see Supplementary Materials
online for complete binding data for the rapamycin experiments).
URL. The latest version of the Java implementation of the GRAM algorithm
may be obtained from the authors’ website at http://psrg.lcs.mit.edu/
GRAM/Index.html.
Note: Supplementary information is available on the Nature Biotechnology website.
ACKNOWLEDGMENTS
Z.B-J. is supported by the Program in Mathematics and Molecular Biology
at Florida State University through the Burroughs Wellcome Fund Interfaces
Program. G.G. is supported by a National Defense Engineering and Science
graduate fellowship. This work was partially funded by a US National Institutes
of Health grant.
COMPETING INTERESTS STATEMENT
The authors declare that they have no competing financial interests.
Received 9 June; accepted 5 August 2003
Published online at http://www.nature.com/naturebiotechnology/
1. Eisen, M.B., Spellman, P.T., Brown, P.O. & Botstein, D. Cluster analysis and display
of genome-wide expression patterns. Proc. Natl. Acad. Sci. USA 95, 14863–14868
(1998).
©
20
03
N
at
ur
e
Pu
bl
is
hi
ng
G
ro
u
p

ht
tp
://
w
w
w.
n
at
ur
e.
co
m
/n
at
ur
eb
io
te
ch
no
lo
gy

L E T T E R S
6 ADVANCE ONLINE PUBLICATION NATURE BIOTECHNOLOGY
2. Segal, E. et al. Module networks: identifying regulatory modules and their condi-
tion-specific regulators from gene expression data. Nat. Genet. 34, 166–176
(2003).
3. Ihmels, J. et al. Revealing modular organization in the yeast transcriptional network.
Nat. Genet. 31, 370–377 (2002).
4. Pilpel, Y., Sudarsanam, P. & Church, G.M. Identifying regulatory networks by combi-
natorial analysis of promoter elements. Nat. Genet. 29, 153–159 (2001).
5. Berman, B.P. et al. Exploiting transcription factor binding site clustering to identify
cis-regulatory modules involved in pattern formation in the Drosophila genome. Proc.
Natl. Acad. Sci. USA 99, 757–762 (2002).
6. Lee, T.I. et al. Transcriptional regulatory networks in Saccharomyces cerevisiae.
Science 298, 799–804 (2002).
7. Mewes, H.W. et al. MIPS: a database for genomes and protein sequences. Nucleic
Acids Res. 30, 31–34 (2002).
8. Holland, M.J., Yokoi, T., Holland, J.P., Myambo, K. & Innis, M.A. The GCR1 gene
encodes a positive transcriptional regulator of the enolase and glyceraldehyde-3-
phosphate dehydrogenase gene families in Saccharomyces cerevisiae. Mol. Cell Biol.
7, 813–820 (1987).
9. Baker, H.V. Glycolytic gene expression in Saccharomyces cerevisiae: nucleotide
sequence of GCR1, null mutants, and evidence for expression. Mol. Cell Biol. 6,
3774–3784 (1986).
10. Forsburg, S.L. & Guarente, L. Identification and characterization of HAP4: a third
component of the CCAAT-bound HAP2/HAP3 heteromer. Genes Dev. 3, 1166–1178
(1989).
11. Matys, V. et al. TRANSFAC: transcriptional regulation, from patterns to profiles.
Nucleic Acids Res. 31, 374–378 (2003).
12. Jacinto, E. & Hall, M.N. Tor signalling in bugs, brain and brawn. Nat. Rev. Mol. Cell
Biol. 4, 117–126 (2003).
13. Crespo, J.L. & Hall, M.N. Elucidating TOR signaling and rapamycin action: lessons
from Saccharomyces cerevisiae. Microbiol. Mol. Biol. Rev. 66, 579–591 (2002).
14. Raught, B., Gingras, A.C. & Sonenberg, N. The target of rapamycin (TOR) proteins.
Proc. Natl. Acad. Sci. USA 98, 7037–7044 (2001).
15. Hardwick, J.S., Kuruvilla, F.G., Tong, J.K., Shamji, A.F. & Schreiber, S.L. Rapamycin-
modulated transcription defines the subset of nutrient-sensitive signaling pathways
directly controlled by the Tor proteins. Proc. Natl. Acad. Sci. USA 96, 14866–14870
(1999).
16. Shamji, A.F., Kuruvilla, F.G. & Schreiber, S.L. Partitioning the transcriptional program
induced by rapamycin among the effectors of the Tor proteins. Curr. Biol. 10,
1574–1581 (2000).
17. Cardenas, M.E., Cutler, N.S., Lorenz, M.C., Di Como, C.J. & Heitman, J. The TOR sig-
naling cascade regulates gene expression in response to nutrients. Genes Dev. 13,
3271–3279 (1999).
18. Hasan, R. et al. The control of the yeast H2O2 response by the Msn2/4 transcription
factors. Mol. Microbiol. 45, 233–241 (2002).
19. Rep, M., Krantz, M., Thevelein, J.M. & Hohmann, S. The transcriptional response of
Saccharomyces cerevisiae to osmotic shock. Hot1p and Msn2p/Msn4p are required
for the induction of subsets of high osmolarity glycerol pathway-dependent genes.
J. Biol. Chem. 275, 8290–8300 (2000).
20. Boy-Marcotte, E., Perrot, M., Bussereau, F., Boucherie, H. & Jacquet, M. Msn2p and
Msn4p control a large number of genes induced at the diauxic transition which are
repressed by cyclic AMP in Saccharomyces cerevisiae. J. Bacteriol. 180, 1044–1052
(1998).
21. Martinez-Pastor, M.T. et al. The Saccharomyces cerevisiae zinc finger proteins Msn2p
and Msn4p are required for transcriptional induction through the stress response ele-
ment (STRE). EMBO J. 15, 2227–2235 (1996).
22. Schuller, H.J. Transcriptional control of nonfermentative metabolism in the yeast
Saccharomyces cerevisiae. Curr. Genet. 43, 139–160 (2003).
23. Crespo, J.L., Powers, T., Fowler, B. & Hall, M.N. The TOR-controlled transcription
activators GLN3, RTG1, and RTG3 are regulated in response to intracellular levels of
glutamine. Proc. Natl. Acad. Sci. USA 99, 6784–6789 (2002).
24. Komeili, A., Wedaman, K.P., O’Shea, E.K. & Powers, T. Mechanism of metabolic con-
trol: target of rapamycin signaling links nitrogen quality to the activity of the Rtg1 and
Rtg3 transcription factors. J. Cell Biol. 151, 863–878 (2000).
25. Liao, X. & Butow, R.A. RTG1 and RTG2: two yeast genes required for a novel path of
communication from mitochondria to the nucleus. Cell 72, 61–71 (1993).
26. Pinkham, J.L. & Guarente, L. Cloning and molecular analysis of the HAP2 locus: a
global regulator of respiratory genes in Saccharomyces cerevisiae. Mol. Cell Biol. 5,
3410–3416 (1985).
27. Dang, V.D., Bohn, C., Bolotin-Fukuhara, M. & Daignan-Fornier, B. The CCAAT box-
binding factor stimulates ammonium assimilation in Saccharomyces cerevisiae,
defining a new cross-pathway regulation between nitrogen and carbon metabolisms.
J. Bacteriol. 178, 1842–1849 (1996).
28. Dang, V.D., Valens, M., Bolotin-Fukuhara, M. & Daignan-Fornier, B. Cloning of the
ASN1 and ASN2 genes encoding asparagine synthetases in Saccharomyces cere-
visiae: differential regulation by the CCAAT-box-binding factor. Mol. Microbiol. 22,
681–692 (1996).
29. Shen-Orr, S.S., Milo, R., Mangan, S. & Alon, U. Network motifs in the transcriptional
regulation network of Escherichia coli. Nat. Genet. 31, 64–68 (2002).
30. Coffman, J.A., Rai, R., Cunningham, T., Svetlov, V. & Cooper, T.G. Gat1p, a GATA
family protein whose production is sensitive to nitrogen catabolite repression, partic-
ipates in transcriptional activation of nitrogen-catabolic genes in Saccharomyces
cerevisiae. Mol. Cell Biol. 16, 847–858 (1996).
©
20
03
N
at
ur
e
Pu
bl
is
hi
ng
G
ro
u
p

ht
tp
://
w
w
w.
n
at
ur
e.
co
m
/n
at
ur
eb
io
te
ch
no
lo
gy

E R R ATA A N D C O R R I G E N D A
Computational discovery of gene modules and regulatory networks
Ziv Bar-Joseph, Georg K Gerber, Tong Ihn Lee, Nicola J Rinaldi, Jane Y Yoo, François Robert, D Benjamin Gordon, Ernest Fraenkel, Tommi
S Jaakkola, Richard A Young & David K Gifford
Nature Biotechnology; doi:10.1038/nbt890
In the version of this article initially published online, the word "and" was omitted from the fourth sentence of the abstract, altering the meaning.
The sentence should read: "We use the GRAM algorithm to describe a genome-wide regulatory network in Saccharomyces cerevisiae using bind-
ing information for 106 transcription factors profiled in rich medium conditions and data from over 500 expression experiments." This mistake
has been corrected for the HTML and print versions of the article.