1. Reconstruction of Transcriptional Regulatory Networks
Adapted from Chapter 4 of
“Systems Biology: Properties of Reconstructed Networks” by
Bernhard O.Palsson

2. What is Transcriptional regulation ?
Transcriptional regulatory networks (TRNs) are the on-off switches at the gene level
Regulating component
Changed RNA and protein output
Changed cell behavior and structures
Input Signals
Adapted from

3. Why do we care about Regulation?
Regulation has a significant effect on cell behavior
Example: E. coli
– Estimated 400 regulatory genes
– 178 regulatory and putative regulatory genes found in genome
– 690 transcription units (contiguous genes with a common expression condition, promoter and terminator) identified in RegulonDB
– Will have a major effect on model predictions of cellular behavior
The basic functional block of a regulatory network is the promoter region of a gene or operon, which contains the cis-regulatory binding sites for the relevant transcription factors regulating the expression of a particular
The TRN is then defined by which TFs bind to which promoters and what the integrated effect of these TFs is on the expression of genes
TRNs can be decomposed into a small set of commonly occuring structural “motifs” and the behavior of prototypical examples of these motifs can be studied to gain insight into their role in the TRN
To begin: why do we care about regulation? When building a mathematical
model it is as important to know what one can neglect in a model as it is to
know what to include. Can we neglect regulation? The answer is, perhaps
under some conditions, but certainly not in all cases. The reason is that
regulation has a significant effect on cell behavior. As an exa mple I have
shown here a table in which E. coli proteins where distributed among 22
functional groups (from the first-draft K-12 annotation). You can see how
metabolism and transport accounts for a substantial number of the known
genes. Additionally, it is estimated that 400 (~10%) of the genes in E. coli
have transcriptional regulatory functions; of these 178 regulatory and putative
regulatory genes have been found in the genome. According to RegulonDB, a
database we will discuss in more detail later, there are 690 transcription units
(e.g. regulated genes or operons) which have been identified in this organism.
Modeling these units will have a major effect on model predictio ns where
regulatory effects have a dominant influence on metabolism. I’ll add in
passing that the effect of regulation is generally much greater in eukaryotes

4. Hierarchy in Transcriptional Regulation
genes
operon
Stimulon
Regulon
For transcriptional regulation there are several levels of abstraction to
consider, and these are some words you will need to know. First of all, you
have the genes. Some genes are constitutive, meaning that they are always
transcribed at relatively constant levels in the cell. Other ge nes are regulated,
as I mentioned, by a stimulus which is sensed, activating a regulator which
either induces or represses transcription of these genes. Sometimes a group of
genes which are neighbors to one another are transcribed together in what is
called an operon. This is primarily found in prokaryotes, presumably because
it is a “cheap” way of regulating several related genes at once. An alternative
means of regulating related genes together is a regulon, where a certain
regulatory protein binds to multiple locations on the DNA, causing induction
or repression (or a combination of both) of related genes or operons. This is
the method of choice for eukaryotes. At the highest level of abstraction is the
stimulon, which includes all of the regulons which are induced by a particular
stimulus.
For example, let’s say that amino acids are present in the extracellular
medium. This is a stimulus which is sensed by the cell, resulting in the
activation or inactivation of certain proteins in the cell. These proteins in turn
have an effect on transcription of various genes and operons, resulting in a new
set of available metabolic or other proteins in the cell. End result: an altered
behavior, most likely in this case that amino acids are no longer synthesized in
the cell de novo, but rather taken up.

5. The lac Operon lac repressor Carbon Source Operation status OFF
Glucose only
CAP
RNA polymerase ON
mRNA
Lactose only
OFF
Neither
OFF
Glucose and Lactose
CAP site
Promoter
Operator
Structural genes for lactose -metobolizing enzymes

6. The GAL Regulon Tup Mig 1 Carbon Source Operation status OFF (basal)
Neither Glucose nor Galactose
Gal4
Gal80
RNA polymerase ON
mRNA
Galactose only
OFF
Glucose and Galactose
Mig 1
Tup
UASG
Mig1 site
GAL1 gene required for galactose metabolism

7. Fundamental data types for TRNs
Component data
Binding sites, transcription factor (TF) molecules etc.
Interaction data
Links are formed by chemical interactions
DNA-protein,protein-protein,metabolite-RNA
Positive and negative controls
Network state data
Reconstructed networks have functional states
Controls for network states assessed by perturbation experiments
Genetic/environmental/systemic

8. Regulatory vs Metabolic Circuits
Regulatory circuits are poorly characterized
Less-well understood
Qualitative statements vs. “hard” stoichiometry
Not mechanistically conserved across different organisms
Regulatory circuits are more complex
Multiple effects/transcription factor (TF)
Multiple regulators/gene
How are regulatory circuits different than metabolic circuits? First, regulatory
circuits are not nearly as well characterized as metabolic circuits. They are
less-well understood and thus far have been characterized more qualitatively in
terms of function (e.g. “protein X has an inhibitory effect on gene abcD ”) in
comparison to metabolic reactions which have quantitative stoichiometric
statements like “Glucose + ATP à Glucose-6-Phosphate + ADP”, which
provides an unambiguous description of the glucokinase catalytic function.
Also, the actual mechanism of many of these circuits is different in different
organisms, and therefore it is more difficult to infer regulatio n from the
genome of an organism as can be done for metabolic genes.
Second, regulatory circuits are more complex than metabolic networks, as
illustrated in this slide of a sea urchin regulatory network. Each transcription
factor can have multiple effects – as described in a regulon – and each gene
may have multiple regulators. All of these factors add to the difficulty of
reconstructing regulatory networks.

9. Key Considerations for TRN Reconstruction
How to represent regulatory information?
– Is transcription regulation Boolean (switch-like) or continuous?
– Should transcription be thought of as a stochastic or
deterministic process?
What constitutes significant regulation?
– Many extracellular signals can affect expression level of a gene.
– Which signals are actually physiologically significant?
Problems with experimental data in the literature:
– Experiments done under different conditions (e.g. strain background)
– Typically experimentalists concentrate on studying well-known TF/target pairs in great detail
In vivo vs in vitro
Because reconstructing regulatory networks in the genome (or in fact any)
scale is such a young field there are still many unresolved issues. A major
issue is how regulatory information should be represented – this is especially
relevant with regards to what type of models are built based on the regulatory
information. In these slides as well as in most of the genetics literature the
assumption is usually made that a gene is either transcribed or not, i.e.
regulation is assumed to be switch like (or Boolean). However, it is well
known by biochemists (but maybe not by geneticists?) that regulatory
processes are in essence no different from other biochemical processes and
that different magnitudes of incoming signals can cause different levels of
transcriptional activity. This issue is illuminated in the following paper:
Biggar SR, Crabtree GR
Cell signaling can direct either binary or graded transcriptional responses.
EMBO J Jun 15;20(12):
Another issue relevant to regulatory reconstruction is what is considered to be
a significant regulatory interaction. In many cases multiple signals can
regulate the transcription of one gene, but some of these signals only play a
modulatory role. These signal are not capable of changing transcriptional
activity on their own and only act in the presence of other stronger signals. A
third problem is how experimental data in the literature should be interpreted.
For example studies on transcriptional regulation in vitro may not be very
useful, because they do not account for other regulatory signals present in vivo.

10. Bottom-up Reconstruction
Pool genomic, biochemical and physiological data, inferring functions where necessary.
include regulatory rules
Represent rules using Boolean logic, kinetic theory and the like.
Analyze separately or together with metabolic network as a metabolic/regulatory model.
Use model to make predictions about the behavior and emergent properties of the system
predictions should be seen as hypotheses which must be tested experimentally.
However, it is possible to reconstruct regulatory networks, given the
information we have, and the process is similar in concept to the process of
metabolic network reconstruction. Again, we need to draw together the
genomic, biochemical and physiological data, inferring functions where
necessary. However, in this case (for bottom- up reconstruction) we will rely
mostly on biochemically characterized regulatory proteins and their
corresponding genes. Rather than including metabolic reactions, we will
include regulatory rules, for example “gene abcD is transcribed if regulatory
protein ProT is active” and “regulatory protein Prot is active if there is oxygen
in the extracellular environment”. These rules can be represented using
Boolean logic, kinetic theory and the like.
The metabolic network may be constructed as usual and now the two networks
may be analyzed separately or together with analytical methods as a
metabolic/regulatory model. Once again, such a model will make predictions
about the behavior and emergent properties of the system which should be
seen as hypotheses which must be tested experimentally.

11. Top-down Reconstruction
Problems with bottom-up reconstruction:
– Many (most?) TF targets are not characterized
– Tedious process, because informative databases are rare
Alternative approach: Utilize data from well-designed high-throughput experiments to reverse-engineer (or “back-calculate”) regulatory circuits
– Gene expression profiles for wild type and deletion strains
under appropriate conditions (genetic perturbation)
– Promoter sequence data and possibly consensus binding sites
for TFs
– Location analysis (ChIP-Chip) data on transcription factor
binding sites
In addition to the issues discussed in the previous slide that are common to any
approach to reconstructing and modeling regulatory networks there are a
couple of specific problems with the bottom- up (literature-based) approach.
The first is due to the fact that up to very recent years targets for transcription
factors were usually identified one gene at a time. While this ensures very low
false positive rates, it also means that most relevant targets for many
transcription factors have not been identified. The second problem is the
general lack of structured databases on transcriptional regulation, which makes
the actual reconstruction process exceedingly time-consuming.
An alternative approach to the bottom- up approach has emerged in the last few
years primarily due to the availability of large-scale gene expression profiling
data. This approach, which we call top-down reconstruction, is based on the
idea that since we know the “outputs” of the complex regulatory circuit in the
form of gene expression profiles, we should be able to reconstruct the
underlying circuit solely from this data. Although gene expression data clearly
is useful for this task it is not the only type of data that could potentially be
utilized. Additional useful data sets are location analysis (or ChIP-Chip) data,
which describes genome-wide binding sites of TFs, and promoter sequence
data (the region upstream of the transcription start site), whic h can be used to
computationally identify binding sites for transcription factors.

12. Issues with Top-down Reconstruction
Very complex models and algorithms are required to reverse engineer regulatory circuits
Computational issues: Explosion in the number of structures
Model complexity issues: Explosion in the number of parameters
Optimality issues: Only locally optimal circuits can be found
Data is not usually available in sufficient quantities or with appropriate quality – computational and experimental people usually don’t work together 
Currently these methods are primarily used to create hypotheses about potential targets of TFs
In addition to the issues mentioned in conjunction with the bottom-down
reconstruction approach, the top-down approach suffers from its own
weaknesses. The major problem is that since the underlying regulatory circuit
is potentially very complex, the types of models and algorithms required for
top-down reconstruction also tend to be very complex (in fact these are
probably some of the most complex statistical models ever constructed). While
this complexity is not a problem as such it results in a few practical problems
in fitting the model that are detailed in the slide. A central problem is the
explosion in the number of different alternative models (both structure and
parameters) to be considered, which requires both large amounts of
sufficiently high-quality experimental data and efficient methods for learning
models from data. Recently, however, some of the best statistical modeling
people have started collaborating with some of the best biology groups that are
capable of generating large quantities of high-quality data. This should lead in
rapid improvement in both the methods for top-down reconstruction and in
new biological knowledge from the reconstruction efforts.

13. Combining knowledge-based and data-based regulatory network reconstruction strategies.a
Regulatory networks can be reconstructed by collecting individual regulatory interactions from relevant databases and the primary literature (knowledge).
Alternatively, networks can be derived directly from high-throughput experimental data and promoter sequence analysis through various data-mining methods.
a Herrgård M.J. et al , Current Opinion in Biotechnology,2004,15:70-77

14. Graphical Representation of Boolean TRNs in E. coli.
Transcription factors (104)
Metabolic genes (479)
Stimuli (102)
TFs regulating gene expression
Stimuli affecting TF activity

15. Summary
Transcriptional regulatory networks determine the expression state of a genome
These networks are presently incompletely defined
Approaches to regulatory reconstruction are still being developed (especially top-down)
Models of TRNs will help unravel the “logic” of gene circuits

16. Thank you !