Defining transcriptional networks through integrative modeling of mRNA expression and transcription factor binding data.

Abstract

Functional genomics studies are yielding information about regulatory processes in the cell at an unprecedented scale. In the yeast S. cerevisiae, DNA microarrays have not only been used to measure the mRNA abundance for all genes under a variety of conditions but also to determine the occupancy of all promoter regions by a large number of transcription factors. The challenge is to extract useful information about the global regulatory network from these data.
We present MA-Networker, an algorithm that combines microarray data for mRNA expression and transcription factor occupancy to define the regulatory network of the cell. Multivariate regression analysis is used to infer the activity of each transcription factor, and the correlation across different conditions between this activity and the mRNA expression of a gene is interpreted as regulatory coupling strength. Applying our method to S. cerevisiae, we find that, on average, 58% of the genes whose promoter region is bound by a transcription factor are true regulatory targets. These results are validated by an analysis of enrichment for functional annotation, response for transcription factor deletion, and over-representation of cis-regulatory motifs. We are able to assign directionality to transcription factors that control divergently transcribed genes sharing the same promoter region. Finally, we identify an intrinsic limitation of transcription factor deletion experiments related to the combinatorial nature of transcriptional control, to which our approach provides an alternative.
Our reliable classification of ChIP positives into functional and non-functional TF targets based on their expression pattern across a wide range of conditions provides a starting point for identifying the unknown sequence features in non-coding DNA that directly or indirectly determine the context dependence of transcription factor action. Complete analysis results are available for browsing or download at http://bussemaker.bio.columbia.edu/papers/MA-Networker/.

Full text

Page 1

BioMed Central
Page 1 of 10
(page number not for citation purposes)
BMC Bioinformatics
Open Access
Research article
Defining transcriptional networks through integrative modeling of
mRNA expression and transcription factor binding data
Feng Gao1, Barrett C Foat1 and Harmen J Bussemaker*1,2
Address: 1Department of Biological Sciences, Columbia University, New York, New York 10027, U.S.A and 2Center for Computational Biology
and Bioinformatics, Columbia University, New York, New York 10032, U.S.A
Email: Feng Gao - fg2037@columbia.edu; Barrett C Foat - bcf2002@columbia.edu; Harmen J Bussemaker* - hjb2004@columbia.edu
* Corresponding author
Abstract
Background: Functional genomics studies are yielding information about regulatory processes in
the cell at an unprecedented scale. In the yeast S. cerevisiae, DNA microarrays have not only been
used to measure the mRNA abundance for all genes under a variety of conditions but also to
determine the occupancy of all promoter regions by a large number of transcription factors. The
challenge is to extract useful information about the global regulatory network from these data.
Results: We present MA-Networker, an algorithm that combines microarray data for mRNA
expression and transcription factor occupancy to define the regulatory network of the cell.
Multivariate regression analysis is used to infer the activity of each transcription factor, and the
correlation across different conditions between this activity and the mRNA expression of a gene
is interpreted as regulatory coupling strength. Applying our method to S. cerevisiae, we find that, on
average, 58% of the genes whose promoter region is bound by a transcription factor are true
regulatory targets. These results are validated by an analysis of enrichment for functional
annotation, response for transcription factor deletion, and over-representation of cis-regulatory
motifs. We are able to assign directionality to transcription factors that control divergently
transcribed genes sharing the same promoter region. Finally, we identify an intrinsic limitation of
transcription factor deletion experiments related to the combinatorial nature of transcriptional
control, to which our approach provides an alternative.
Conclusion: Our reliable classification of ChIP positives into functional and non-functional TF
targets based on their expression pattern across a wide range of conditions provides a starting
point for identifying the unknown sequence features in non-coding DNA that directly or indirectly
determine the context dependence of transcription factor action. Complete analysis results are
available for browsing or download at http://bussemaker.bio.columbia.edu/papers/MA-Networker/
.
Background
For various organisms, DNA microarrays have been used
to measure the mRNA abundance for essentially all pro-
tein-coding genes in the genome under a large number of
conditions [1,2]. Microarray technology can also be com-
bined with chromatin-immunoprecipitation (ChIP) or
chromatin profiling (DamID) to quantify the occupancy
of upstream non-coding regions by transcription factors
or other chromatin-associated proteins [3-9]. In the bud-
ding yeast Saccharomyces cerevisiae, ChIP has been used to
Published: 18 March 2004
BMC Bioinformatics 2004, 5:31
Received: 12 November 2003
Accepted: 18 March 2004
This article is available from: http://www.biomedcentral.com/1471-2105/5/31
© 2004 Gao et al; licensee BioMed Central Ltd. This is an Open Access article: verbatim copying and redistribution of this article are permitted in all media
for any purpose, provided this notice is preserved along with the article's original URL.

Page 2

BMC Bioinformatics 2004, 5http://www.biomedcentral.com/1471-2105/5/31
Page 2 of 10
(page number not for citation purposes)
globally map the binding sites of over a hundred tran-
scription factors [4]. Moreover, mRNA expression data for
over a thousand conditions has been published. The chal-
lenge is to find new ways to extract knowledge about the
regulatory mechanisms that govern the cell by combining
these different types of data [10-14].
Initiation of transcription in eukaryotes is a complicated
process that depends on the binding of transcription fac-
tors (TFs) and chromatin-modifying enzymes to the pro-
moter region as well as the recruitment of the RNA
Polymerase II complex to the transcription start site. Tran-
scriptional control is combinatorial, and cooperative
binding of multiple factors on the same promoter region
and/or cooperative recruitment of the Pol II complex is
often required for transcriptional activation [15]. Occu-
pancy of the promoter region of a gene by a transcription
factor is thus a necessary but not a sufficient condition for
the gene to be controlled by it. As a consequence, genome-
wide transcription factor binding patterns measured using
ChIP or DamID microarray experiments alone can only
indicate the potential for a gene to be regulated by a given
TF. Independent information will be required to establish
that the gene is indeed a functional target of the factor.
Deletion or over-expression of a transcription factor, com-
bined with genomewide microarray profiling of the differ-
ence in expression between mutant and wild type, is also
widely used to infer regulatory interactions. However,
drastic perturbation of the genetic network outside the
physiologically relevant range may lead to false target pre-
diction, or the mutant strain may simply not be viable.
Moreover, direct and indirect targets of the factor cannot
be distinguished using this approach.
When mRNA abundances for all genes are compared
between two experimental conditions in a microarray
experiment, the observed differential expression pattern is
usually a superposition of responses of various pathways,
mediated by signaling cascades that end at the level of
transcription factors. It has previously been shown that
these changes in TF activity can be quantitatively inferred
by performing multivariate regression analysis on the
expression log-ratios from a single microarray experiment
[16-19]. Transcription factors are implicitly represented
by a consensus motif for their DNA binding sites, and the
regression coefficients estimate the changes in TF activity.
In the present study we will instead use ChIP occupancy
log-ratios as predictors for expression. No sequence infor-
mation will be used, and it is therefore not necessary that
a DNA consensus motif be known for the TFs. Again, mul-
tivariate regression analysis of a single genomewide set of
mRNA log-ratios on the genomewide binding profiles of
a large number of TFs for which ChIP data is available can
be used to quantify to what extent each transcription fac-
tor is responsible for the observed changes in mRNA
expression.
When the regression procedure is performed in parallel on
a large library of expression data, the inferred changes in
TF activity for each comparison can be combined into a
transcription factor activity profile (TFAP) for each tran-
scription factor ("Step 1" in Fig. 1a). Each TFAP represents
a highly specific regulatory signature, as is shown for three
transcription factors in Fig. 1b: The activity of the G2
phase related factor Ndd1p oscillates during the cell cycle
(blue), but shows little or no response to nutrient deple-
tion and other stress conditions (red), or changes in alpha
pheromone concentration (green). Complementary
behavior is seen for the TCA cycle regulator Hap4p and
the mating-related factor Ste12p.
One expects the mRNA expression profile of a gene regu-
lated by a specific transcription factor to be similar to the
TFAP of that factor. We therefore investigated whether the
linear correlation across the experiment library between a
TFAP and the mRNA expression profile of a gene whose
promoter is bound by the factor could be interpreted as a
regulatory coupling strength and used to improve the spe-
cificity of target prediction. To this end, we constructed a
matrix of regulatory coupling strengths between all tran-
scription factors and all genes ("Step 2" in Fig. 1a). When
this information is combined with the original ChIP data
for a given TF, the ChIP log-ratio and coupling strength for
each gene can be shown simultaneously in a 2D scatter
plot (Fig. 1c). The fact that each gene has two parameters
associated with it allows a more sophisticated classifica-
tion than is possible based on ChIP alone. We first
defined a set "B+" of genes that are significantly bound by
a TF (we required the P-value reported by Lee et al. to be
smaller than 10-3) [4]. Next, we then partitioned the "B+"
gene set into two subsets "B+/C+" and "B+/C-" based on
whether or not their mRNA expression profile was signif-
icantly correlated with the TFAP (Pearson correlation, 5%
false discovery rate). Our hypothesis is that the B+/C+
genes (shown in red in Fig. 1c) are the functional direct
targets of the factor, while the binding to the promoter
region of the B+/C- genes (shown in green) is non-func-
tional.
Results and discussion
Only a subset of genes bound by each TF is controlled by it
We have focused on the yeast Saccharomyces cerevisiae, for
which a wealth of functional genomics data is available.
We compiled a library of ~750 expression patterns from
various sources, and combined it with the genome-wide
promoter occupancies in mid-log phase and rich medium
for 113 TFs as measured by Lee et al. [4]. It was determined
that 37 transcription factor occupancy patterns out of the

Page 3

BMC Bioinformatics 2004, 5http://www.biomedcentral.com/1471-2105/5/31
Page 3 of 10
(page number not for citation purposes)
(a) Overview of our method for determining regulatory coupling strengths between transcription factors and their putative target genes
Figure 1
(a) Overview of our method for determining regulatory coupling strengths between transcription factors and their putative
target genes. Inputs are (i) a library of microarray expression data for a large number of conditions and (ii) genomewide (ChIP)
occupancy data for one or more transcription factors. In the first step of our algorithm, a matrix of transcription factor activi-
ties is inferred by using regression analysis to explain the mRNA expression pattern under each condition in terms of the ChIP
data for each transcription factor. In the second step, a matrix of regulatory coupling strengths is determined by computing the
correlation between each transcription factor activity profile (TFAP) and the mRNA expression profile of each gene. (b)
Examples of transcription factor activity profiles. The activity profiles of three transcription factors (Hap4, Ndd1, Ste12) are
shown across stress response, pheromone response, and cell cycle [28-30]. Significant changes in activity of the TCA cycle reg-
ulator Hap4p occur mostly in metabolic stress conditions, while changes in the activity of the cell cycle regulator Ndd1p and
the pheromone-dependent regulator Ste12p are associated with the cell cycle and signal transduction experiments, respec-
tively. (c) Examples of scatter plots of ChIP binding log-ratio versus coupling factor. In the scatter plots, black dots denote
unbound (B-) genes, red dots denote bound and coupled genes (B+/C+), while green dots denote genes that are bound but not
coupled (B+/C-). A threshold of P = 10-3 was used to determine significant binding as described in Lee et al. [4]. A threshold for
coupling was determined by requiring a false discovery rate of 5%, as described in Methods.

Page 4

BMC Bioinformatics 2004, 5 http://www.biomedcentral.com/1471-2105/5/31
Page 4 of 10
(page number not for citation purposes)
full set of 113 are significant predictors of mRNA expres-
sion for one or more experiments in our library (see Meth-
ods). Note that the library of expression data we used
obviously does not cover all possible experimental condi-
tions and our method is therefore likely to underestimate
the number of transcription factors that are present in the
nucleus under the conditions used by Lee et al. [4]. For
each of the 37 factors selected for further analysis, a tran-
scription factor activity profile (TFAP) was computed
("Step 1" in Fig. 1a) and the TF-target coupling strength
was determined ("Step 2" in Fig. 1a). The number of genes
in each group (B-, B+/C+, and B+/C-) is listed in Table 1
for the 37 transcription factors analyzed. On average 58%
of significantly bound genes are classified as significantly
coupling genes. Activity profiles and B+/C+ target predic-
tions for all 37 factors are available on the website sup-
porting this paper.
Enrichment for specific functional categories
Several analyses were performed to validate our results.
First, we established that B+/C+ genes are significantly
enriched for specific Gene Ontology (GO) categories
(hypergeometric distribution; 5% false discovery rate)
[20]. This result is not surprising, as we would already
expect the set B+ of ChIP positives per se to be enriched for
roughly the same functional categories. By contrast, for
Table 1: Classification of genes according to ChIP data and inferred regulatory coupling.
TFB- B+/C+B+/C-Unclassified*
Abf1
Ace2
Arg81
Bas1
Cad1
Dal81
Dig1
Fhl1
Fkh2
Gal4
Gat3
Gcn4
Hal9
Hap4
Hir1
Hir2
Hsf1
Leu3
Mbp1
Mcm1
Met31
Msn4
Mss11
Ndd1
Nrg1
Rlm1
Sig1
Sko1
Sok2
Stb1
Ste12
Sum1
Swi4
Swi5
Thi2
Yap1
Yap6
5638
5843
5985
5975
5854
5823
5872
5754
5261
5149
5891
5919
5948
5939
5963
5932
5929
5988
5641
5709
5983
5952
4735
5799
5912
5900
5683
5155
4881
5791
5725
5919
5528
5358
5940
5959
5924
138
33
11
28
26
24
20
146
61
20
40
58
4
47
19
8
35
8
65
42
15
23
11
50
61
13
0
2
10
13
19
35
75
47
5
28
33
136
33
10
13
13
16
11
37
45
20
29
21
13
21
10
13
17
13
40
46
13
7
9
42
20
24
0
0
6
10
31
26
48
45
1
16
54
470
473
376
366
489
519
479
445
1015
1193
422
384
417
375
390
429
401
373
636
585
371
400
1627
491
389
445
699
1225
1485
568
607
402
731
932
436
379
371
*Genes whose expression level in microarray data or binding P-value in Lee et al. is not available. The number of genes in each of the categories (B-
, unbound genes; B+/C+, bound and coupling genes; B+/C-, genes that are bound but do not couple) is shown for each of the 37 transcription
factors analyzed. On average, 58% of the significantly bound genes were classified in the B+/C+ group.

Page 5

BMC Bioinformatics 2004, 5http://www.biomedcentral.com/1471-2105/5/31
Page 5 of 10
(page number not for citation purposes)
almost all TFs analyzed we found no significant enrich-
ment for any GO category in the set of non-coupling (B+/
C-) genes (Fig. 2, see supplementary website for details).
This result is very significant because it suggests that our
criterion for distinguishing functional from non-func-
tional TF targets based on regulatory coupling is accurate:
There seems to have been no evolutionary pressure on the
set of B+/C- genes.
Transcriptional response to transcription factor deletion
Next, we analyzed the expression response to transcrip-
tion factor deletion for the B+/C+ and B+/C- genes. The
mean and the standard deviation of the gene expression
log-ratio between mutant and wild type as obtained in
Hughes et al. were calculated for all genes in the genome,
as well as for the B+/C+ and B+/C- groups [21]. A sample
t-test was performed to determine the significance of the
change in expression. The B+/C+ genes show a significant
change in mRNA expression for the 7 transcription factors
for which deletion and ChIP data are both available. By
Enrichment for functional annotation
Figure 2
Enrichment for functional annotation. The number of significantly over-represented Gene Ontology (GO) categories in the
group B+/C+ of genes that couple to transcription factor activity (red) and the group B+/C- of genes that do not couple
(green) for each of the 37 transcription factors analyzed. No significant enrichment for any GO category was found in most B+/
C- gene groups, supporting the hypothesis that only the coupling B+/C+ genes are functional targets.
Table 2: Analysis of transcriptional response to transcription factor deletion.
Transcription
Factor
Genomewide B+/C+B+/C-
MeanSD Mean SD-log10(p) MeanSD -log10(p)
Dig1
Gcn4
Hir2
Mbp1
Swi4
Swi5
Yap1
0.759
0.785
0.724
0.830
0.524
0.744
0.756
0.198
0.177
0.218
0.121
0.338
0.204
0.201
0.567
0.294
0.180
0.541
0.348
0.583
0.471
0.360
0.329
0.295
0.324
0.352
0.373
0.376
3.76
28.23
4.00
27.86
4.89
5.98
7.66
0.846
0.793
0.760
0.770
0.533
0.759
0.587
0.080
0.170
0.194
0.172
0.333
0.186
0.337
0.04
0.24
0.14
2.79
0.24
0.17
2.67
The mean and the standard deviation of gene expression log-ratio between mutant and wild type as obtained in Hughes et al. were calculated for all
genes in the genome as well as for the B+/C+ and B+/C- groups [21]. A sample t-test was performed to determine the significance of the change in
expression. The B+/C+ genes show a significant change in mRNA expression for the 7 transcription factors for which deletion and ChIP data is
available. By contrast, the response of the B+/C- genes is insignificant for most transcription factors.