Reverse engineering of genetic networks
with time delayed recurrent neural networks
and clustering techniques
Dissertation
submitted to the
Combined Faculties
for the Natural Sciences and for Mathematics
of the Ruperto-Carola University of Heidelberg, Germany
for the degree of
Doctor of Natural Sciences
presented by
M. Sc. David Camacho Trujillo
born in México City, México
Oral-examination: ................................................
2
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
.............................................................
..................................................
..................................................
Referees:
.
Prof. Dr. Ursula Kummer
P.D. Dr. Ursula Klingmüller
3
4
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
D edicated to:
Sarah
&
Tere
&
A rturito
5
6
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
7
INDEX
Summary.......................................................................... 9
Zusammenfassung ....................................................... 10
Personal Words............................................................. 11
List of abbreviations ..................................................... 13
General Motivation........................................................ 17
1. Biological context ..................................................... 19
1.1 Gene regulation ............................................................................................. 19
1.2 Basal transcription apparatus ......................................................................... 19
1.3 Transcription factors...................................................................................... 21
1.4 Enhancers-Insulators...................................................................................... 22
1.5 Post-transcriptional regulation of the mRNA ................................................. 23
1.5.1 Alternative splicing................................................................................. 23
1.5.2 RNA interference.................................................................................... 24
1.5.3 Dimensional in-homogeneities ................................................................ 26
2. Reverse engineering and modelling of genetic
network modules........................................................... 29
2.1 Related work ................................................................................................. 29
2.2 General concepts ........................................................................................... 30
2.3 Dimensionality reduction by data selection .................................................... 32
2.4 Theoretical works .......................................................................................... 36
2.4.1 Boolean Networks................................................................................... 36
2.4.2 Differential equation systems .................................................................. 38
2.4.3 Stochastic Models ................................................................................... 44
2.4.4 Bayesian networks .................................................................................. 45
3. Methods ..................................................................... 50
3.1 Workflow ...................................................................................................... 50
3.2 Data pre-processing, Quality control.............................................................. 51
3.3 Data normalization ........................................................................................ 53
8
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
3.4 Dimensionality problem. The use of interpolation approaches ....................... 55
3.5 Data fitting .................................................................................................... 57
3.6 Models........................................................................................................... 62
3.6.1 The CTRNN model................................................................................. 62
3.6.2 The TDRNN model................................................................................. 66
3.6.3 Robust parameter determination.............................................................. 67
3.6.4 Graph generation and error distance measurements ................................. 68
3.6.5 Clustering of results ................................................................................ 68
3.6.6 Dynamic Bayesian Network.................................................................... 71
4. Results ....................................................................... 73
4.1 Synthetic benchmark: The Repressilator ........................................................ 74
4.1.1 Parameter space selection........................................................................ 75
4.1.2 Required data length. .............................................................................. 86
4.1.3 Robustness against noise......................................................................... 92
4.1.4 Robustness against incomplete information: Clustering improves the
standard reverse engineering task, quantitatively and qualitatively................... 97
4.2 The yeast cell cycle...................................................................................... 103
4.2.1 TDRNN shows superior inference and predictive power than previous
models on experimental data.......................................................................... 104
4.2.2 Bootstrapping validation ....................................................................... 106
4.2.3 Clustering improves the RE process with real data ................................ 107
4.3 Reverse engineering of keratinocyte-fibroblast communication.................... 109
5. Discussion ............................................................... 127
5.1 Model choice and data driven experiments................................................... 128
5.2 Data selection .............................................................................................. 129
5.3 Data interpolation, implications ................................................................... 130
5.4 Data fitting and inference power relationship ............................................... 131
5.5 Reverse engineering framework, improving the robust parameter selection.. 135
6. Conclusions............................................................. 137
7. Bibliography ............................................................ 139
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
9
Summary
In the iterative process of experimentally probing biological networks and
computationally inferring models for the networks, fast, accurate and flexible
computational frameworks are needed for modeling and reverse engineering
biological networks. In this dissertation, I propose a novel model to simulate gene
regulatory networks using a specific type of time delayed recurrent neural networks.
Also, I introduce a parameter clustering method to select groups of parameter sets
from the simulations representing biologically reasonable networks. Additionally, a
general purpose adaptive function is used here to decrease and study the connectivity
of small gene regulatory networks modules.
In this dissertation, the performance of this novel model is shown to simulate the
dynamics and to infer the topology of gene regulatory networks derived from
synthetic and experimental time series gene expression data. Here, I assess the quality
of the inferred networks by the use of graph edit distance measurements in
comparison to the synthetic and experimental benchmarks. Additionally, I compare
between edition costs of the inferred networks obtained with the time delay recurrent
networks and other previously described reverse engineering methods based on
continuous time recurrent neural and dynamic Bayesian networks. Furthermore, I
address questions of network connectivity and correlation between data fitting and
inference power by simulating common experimental limitations of the reverse
engineering process as incomplete and highly noisy data.
The novel specific type of time delay recurrent neural networks model in combination
with parameter clustering substantially improves the inference power of reverse
engineered networks. Additionally, some suggestions for future improvements are
discussed, particularly under the data driven perspective as the solution for modeling
complex biological systems.
10
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Zusammenfassung
Für den iterativen Prozess der experimentellen Erforschung biologischer Netzwerke
und der computergenerierten Ableitung von Modellen für diese Netzwerke werden
schnelle, fehlerfreie und flexible Programmiergerüste benötigt, um biologische
Netzwerke zu modellieren und um sie zu rekonstruieren. In dieser Arbeit stelle ich ein
neuartiges Modell vor, das genregulierte Netzwerke darstellt, indem zeitverzögerte,
rekurrente, neuronale Netzwerke benutzt werden. Zudem führe ich eine Methode des
Parameter-Clusterings ein, die Parameter-Set-Gruppen, die biologisch sinnvolle
Lösungen darstellen, aus den Simulationen auswählt. Zusätzlich wird hier eine
generelle, lernfähige Funktion eingesetzt, um die Konnektivität kleiner genregulierter
Netzwerke zu verringern und um diese zu untersuchen.
In dieser Dissertation wird die Leistungsfähigkeit dieses neuartigen Modells, die
Dynamik
genregulierter
Netzwerke
aus
synthetischen
und
experimentellen
Datensätzen von Zeitreihen der Gen-Expression zu simulieren und deren Topologie
abzuleiten, aufgezeigt. Die Qualität der abgeleiteten Netzwerke bestimme ich mit
Hilfe von Graph-Edit-Messungen im Vergleich zu den synthetischen und
experimentellen Bezugswerten. Außerdem vergleiche ich den Arbeitsaufwand der von
den zeitverzögerten rekurrenten Netzwerken abgeleiteten Netzwerke und anderer
bereits beschriebener Rekonstruktionsmethoden, die auf zeitkontinuierlichenrekurrenten und dynamischen-bayesischen Netzwerken basieren. Darüber hinaus
befasse ich mich mit Fragen der Netzwerk-Konnektivität und der Korrelation
zwischen der Datenanpassung und der statistischen Power der Inferenz, indem ich
bekannte experimentelle Einschränkungen des Rekonstruktionsprozesses, wie
unvollständige oder höchst rauschbehaftete Datensätze, simuliere.
Dieses neuartige und spezielle, zeitverzögerte, rekurrente, neuronale Netzwerk
verbessert zusammen mit dem Parameter-Clustering wesentlich die Ableitungskraft
der rekonstruierten Netzwerke. Zudem werden einige Anregungen für zukünftige
Verbesserungen erörtert, insbesondere aus der datengestützen Perspektive als der
Lösungsstrategie für die Modellierung komplexer biologischer Systeme.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
11
Personal Words
I would like to acknowledge in this dissertation to those persons and institutions that
made this work possible.
The support of Professor Ursula Kummer is of special importance due to the
circumstances of this work. Therefore, I would like to sincerely acknowledge her for
this opportunity to finish all these years of work on a nice way.
I want to acknowledge Professor Randall Beer, who thanks to the internet could
provide me with punctual but decisive directions to understand and develop my
theoretical model. Analogous are the contributions from Professor Paul Johnson as
well as the SWARM community who always helped me to implement my model
under the SWARM philosophy. Additionally, I would like to thanks Dr. J.J. Merello
for his GA code and directions to use it.
Undoubtedly, this work could not be possible without the support from Sarah. In
many ways she has supported me and was at my side in difficult moments. Hence, I
want to set this in words. Many thanks for all your support, Saritah.
Special thanks to my family because despite they are not geographically close to me,
they always were present and support me with their comprehension. Special thanks to
my sisters Rebekita and Lendy, for their tenderness and shift responses when needed.
I acknowledge the opportunity to develop this work at the iBIOS group from the
Deutsches Krebsforschungszentrum (DKFZ) and the Viroquant group at Bioquant
from the University of Heidelberg.
This work could be developed by the support of the DAAD who I would like to
acknowledge, because it is more than a building - they always have been a human
Institution.
In this sense, I would like to acknowledge my former University UNAM, because
despite of being massive, it has a quality similar to the University of Heidelberg and
gave me my formation for free. Once I heard that that massive universe imprint over
us some sort of nationality, and I believe there is a little truth in that.
I would like to thank all my friends that have been at my side, in one or another way.
Finally, I would like to thanks to my own “dickopfness” because in many senses, it
was not planned for me to be at this point.
12
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
List of abbreviations
Activating Protein 1
AP-1
Adaptive time-delay neural network
ATNN
ATP: protein phosphotransferase (cAMP-dependent)
PKA
Boolean network
BN
Carcinoembryonic antigen-related cell adhesion molecule 1
CEACAM1
Catalytic subunit of the main cell cycle - cyclin-dependent kinase CDK (yeast)
cdc28
Cell division control protein 15 (yeast)
cdc15
Cell division cycle 14 protein (yeast)
Cdc14
Cell division cycle 20 protein (yeast)
Cdc20
Cell division cycle 20-like protein 1(yeast)
Cdh1
Central nervous system
CNS
Connectivity
K
Continuous time recurrent neural network
CTRNN
Cyclin-dependent kinase (yeast)
Mcm1
Deoxyribonucleic Acid
DNA
Dynamic Bayesian network
DBN
Early growth response protein 1
EGR1
Epidermal grow factor receptor
EGF-R
Epidermal grow factor
EGF
Escherichia Coli
E.coli
FBJ murine osteosarcoma viral oncogene homolog
FOS
Fetal calf serum
FCS
G1/S-specific cyclin (yeast)
Cln 3
G1/S-specific cyclins (yeast)
Cln 1/2
G2/mitotic-specific cyclin 1/2(yeast)
Clb 1/2
Gene regulatory network
GRN
Gene Set Enrichment Analysis
GSEA
Genetic algorithm
GA
Glyceraldehyde-3-phosphate dehydrogenase (NADP+)
GAPDH
Granulocyte–macrophage colony-stimulating factor
GM-CSF
13
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
14
Graph edit distance
GED
Gray
Gy
(absorbed radiation units)
Hepatocyte growth factor
HGF
Human skin dermal fibroblasts
HDF
Immortalized human keratinocytes
HaCaT
Integrin beta 6
ITGB6
Integrin, alpha V
ITGAV
Irradiated human skin dermal fibroblasts
HDFi
Keratinocyte growth factor
FGF-7
Laminin, alpha 3
LAMA3
Laminin, gamma 2
LAMC2
Means square error
MSE
Messenger ribonucleic acid
mRNA
Micro ribonucleic acid
miRNA
Mismatch
MM
Multiple recurrent neural networks
MRNN
Multiprotein bridging factor (yeast)
MBF
Nondeterministic polynomial-time hard
NP-hard
Not pruning
NP
Open reading frame
ORF
Ordinary differential equations
ODE
Perfect match
PM
Plasminogen activator, urokinase
PLAU
Plasminogen activator, urokinase receptor
PLAUR
Prostaglandin-endoperoxide synthase 2
PTGS-2
Pruning
P
Random Boolean networks
RBN
Regulatory protein SWI4
(yeast)
SBF
Regulatory protein SWI5
(yeast)
Swi5
Reverse engineering
RE
Reverse transcriptase-polymerase chain reaction
RT-PCR
Ribonucleic acid
RNA
Ribonucleic acid polymerase subunit II
RNA pol II
Simulating annealing
SA
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
S-phase entry cyclin 5/6
(yeast)
Stromal derived factor-1
Clb5/6
SDF-1
Substrate and inhibitor of the cyclin-dependent protein
kinase CDC28
(yeast)
Sic1
Thiamine-adenine promoter consensus sequences
TATA box
Time delayed neural networks
TDNN
Time delayed recurrent neural network
TDRNN
Transcription factors
TF
Transcription factor Jun B
JUNB
Variance stabilization normalization
VSN
v-ets erythroblastosis virus E26 oncogene homolog 1
ETS1
15
16
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
17
General Motivation
A group of technologies like microarrays, CGH or mass spectrometry, has become
part of the standard laboratory experiments all around the word. These technologies
allow us to measure thousands of genes, hundreds of proteins or other cellular
components like mRNA´s at the same time. All this information is usually primary
stored into databases but generally it analysis is far to be finished. One reason for this
situation is that the traditional one to one “cause-effect” correlation, typically used in
biology, is not applicable to these large data sets. To handle this information a new
kind of approaches has been developed during the last years. Approaches able to
store, analyze and develop models for a large number of variables.
One area that is deeply influenced by experimental high throughput data generation
technologies is the analysis of gene regulation of the mammalian cells. Because of its
complexity and implications in different areas like evolution or drug target
generation, gene regulation is widely studied by theoretical works. The holistic
integration of gene regulation dynamics has just begun, and it is clear that only the
iterative work between lab data driven experiments and theoretical work will be able
to generate a new paradigm in the area. In this multidisciplinary context the present
work is circumscribed. To understand the goals, achievements and limitations of this
work, some basics topics will be described about gene regulation complexity (Chapter
1), and the most relevant related theoretical work developed until now (Chapter 2). In
Chapter 3 the methodology used in this thesis is described. A comparative study to
with similar approaches is achieved in the results section (Chapter 4) as well as the
presentation of results obtained by applying the approach described in the present
work, to experimental data. Lastly, the analyses of the results as well as collateral
topics are described on the discussion section (Chapter 5) and some final words and
outlook is described in the Conclusion section (Chapter 6).
18
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
19
1. Biological context
1.1 Gene regulation
Gene regulation is a complex and not well-understood process. It has several
mechanisms to control itself which act at different levels in time scale, cellular space
and molecular mechanisms. Here, some of these mechanisms will be described to
highlight the implications and restrictions they impose on the theoretical models
intended to capture gene regulation behavior.
1.2 Basal transcription apparatus
The basal transcription apparatus is not part of the gene regulation mechanism by
itself. However, it is important to remark that its presence is a necessary condition in
order to transcribe any gene. Therefore, it is important to know some structural
aspects of it, which plays a role for the design of gene regulation kinetic related
models (Mjolsness and Sharp, 1991).
While the enzymatic behavior of transcription is due to the basal transcription
apparatus bound to the tetrameric RNA polymerase II, substrates are the relative free
diffusible nucleotide bases and the highly conserved thiamine-adenine promoter
sequences (TATA boxes, here on) on the DNA. Finally, the mRNA is the obtained
product. From here on, it should be clear that the TATA boxes do not form part of the
same chemical liquid phase as the other substrates. TATA boxes are part of an
extreme long polymer associated to thousands other proteins conforming a dynamical
semi-solid phase system, the so-called Chromatin. Therefore, traditional kinetic
models (as Michaelis Menten and others) should not be applicable here, because they
presuppose a freely diffusive Brownian motion of all substrates along an
homogeneous liquid phase media, meaning homogeneous concentration.
20
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
The basal transcription apparatus plus the tetrameric RNA pol II is a complex of more
than 20 proteins that is in situ sequentially assembled This multimeric complex
requires additional proteins to initiate its formation; the so-called Mediator (Lewis
and Reinberg, 2003) is part of those required additional proteins. The Mediator is
about 20 proteins in size and this large multimeric protein complex requires additional
proteins to initiate the transcription, the so-called Transcription Factors (TF here on).
These TF are the triggering initial step in gene transcription activation.
The basal transcription apparatus plus Mediator and additional TF is a large
multimeric complex, which could have the size of more than 60 proteins. Hence, it is
clear that gene transcription initiation requires some structural conditions on the DNA
super structure as accessibility to avoid steric impediments. This accessibility is
controlled by other regulatory mechanisms that promote chromatin relaxation
(Cremer and Cremer, 2001). Additional aspects like DNA malleability (3D curvature
of the DNA) or stochastic fluctuation in access generated by the Chromatin
“breathing” are a matter of discussion. For sure, these processes play also their roll,
but the problem is to know when and how intense every regulatory mechanism
contributes to the global behavior. From here on, it should be clear that chromatin
accessibility is a tri dimensional level of inhomogeneity and a source of gene activity
regulation. Models oriented to represent gene reaction-diffusion kinetics should take
this into account.
Concerning the product (mRNA) kinetics; since the mRNA is a polymer, it follows a
multiple-step-synthesis. Additionally, as this mRNA product is not a monotonic
polymer, its step-by-step formation implies a more complex process (among other
processes like translocation, strand separation etc.) known as nucleotide selection.
This means that the rate of the mRNA production is not diffusion limited, and has
slight variations depending on the template sequence and other context dependent
proteins. On the other hand, once the basal transcription machinery is induced it
activates the RNA polymerase, and in turn it moves along the gene performing the
transcription. The consequence of this RNA polymerase displacement is that another
RNA polymerase is able to bind the initiation complex and initiate simultaneously a
new transcription process.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
21
Therefore, even though there is only one functional a copy of every gene, it is difficult
to quantify the maximum number of transcription complexes at a certain time. Owing
to these last two considerations, it is practically impossible to define a transcription
rate constant for any gene, because the rate of transcription for every gene is a
continuous time-and-context dependent process.
1.3 Transcription factors
Transcription factors belong to a large but limited set of families of proteins that share
functionality, in human they are about 300 (Itzkovitz, et al., 2006) TF can recognize
(the mechanism is family specific) (Itzkovitz, et al., 2006) pattern sequences into the
DNA along the so-called promoter region of every gene. Once the TF are bound to
the promoter region of a given gene, they can interact with proteins from the basal
transcription apparatus and together recruit the RNA polymerase II initiating the
transcription of that gene. The activation strength is function of the concentration and
physical interactions among the TF that activate a given gene at a certain time.
It has been shown (van Nimwegen, 2003) that the total number of TF (N) of any
specie scales with it genome size (G) as a power-law (N∼G1.9 Prokaryotes, N∼G1.3
Eukaryotes). However, all of them are not present at the same time nor with the
required concentration to activate their target genes. Instead, TF need to interact with
some others proteins in order to activate one gene. Usually, they form dimers (homo
and hetero dimers) and in turn form quaternary complexes at the promoter regions
(Pilpel and Sudarsanam, 2001), where typically more than one complex regulates its
activity. Furthermore, it has been proposed that the total number of TF per family
correlates with the number of degrees of freedom (number of base pairs recognized by
family, ranging from 4 to 96 in humans) in the binding mechanism. However, an
overlap of sequence recognition occurs among different TF, probably to make the
system more robust through redundancy.
Regarding the concentration of the TF in the nucleus, the general idea is that the local
concentrations of TF are responsible for the activation of their target genes.
22
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
The local variation of TF concentrations is influenced by their transport
(internalization into the nucleus from the cytoplasm) which in turn is often regulated
by their activation (usually by phosphorylation) and negatively regulated by their
deactivation and or degradation. However, once the TF are in the nucleus, local
variations of TF concentrations occur due to the interaction with other already
focalized (attached to) proteins at certain DNA regions, the so-called enhancers.
Interactions between TF and its promoter targets on the DNA are not covalent, the
real picture is a dynamical stochastically process where TF are bound and unbound
permanently to the DNA. In this sense TF activation state often plays a central role in
their activity because often the activated (usually phosphorylated) transcription
factors exhibit a higher affinity for the DNA recognition site, but when the
concentration of inactive TF is high enough, then the none activated TF could
displace the already bound active TF. Therefore, at this TF regulation level, gene
activity again is a continuous time combinatorial process, function of TF identities,
their transport, local concentrations and often it activation state (phosphorylated or
not).
1.4 Enhancers-Insulators
These are short regions on the DNA, which can facilitate the transcription of (cis)
genes at a relative long-distance. They are defined like distant-acting cis-regulatory
elements (Blackwood and Kadonaga, 1998) but their precise mechanism of acting is
still not clear. There is evidence that enhancers increase the probability of genes on
their surroundings to be transcribed. Enhancers do their task probably by increasing
the local concentration (known as nuclear localization) of TF, but there are some
other proposed mechanisms such as; chromatin or nucleosomes remodeling,
superhelical tension (to facilitate chromatin accessibility) and direct interaction with
associated proteins and the transcription basal machinery. Enhancers also have their
functional counterpart on the so-called insulators (West, et al., 2002), which probably
directly inhibit the functioning of enhancers, but probably they promote gene
repression by other mechanism like Chromatin condensation.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
23
However, there is not enough information on the final balance between enhancersinsulators until now. Therefore this could be modeled like a none-specific Bias of the
gene activation process.
1.5 Post-transcriptional regulation of the mRNA
At this point two mechanisms are the most relevant to take into account for modeling
gene regulatory networks: alternative splicing and interference RNA. A good
description of both could be found elsewhere therefore here the focus is just on the
implications for the area of modeling gene regulatory networks.
1.5.1 Alternative splicing
The mRNA is transcribed as a precursor containing intervening sequences (introns).
These sequences are subsequently removed such that the flanking regions (exons) are
spliced together to form mature mRNA. Alternative splicing pathways generate
different mRNAs encoding distinct protein products, those increasing the coding
capacity of genes. The resulting proteins may exhibit different and sometimes
antagonistic (Cremer and Cremer, 2001) functional and structural properties, as
binding affinity, intracellular localization, enzymatic activity, stability and posttranslational modifications, and may inhibit the same cell with the resulting
phenotype being the balance between their expression levels. Alternative splicing can
also act as an on–off gene expression switch by the introduction of premature stop
codons (Feyzi, et al., 2007).
The alternative splicing mechanism is achieved by a ribonucleo-protein structure
called spliceosome, but most of the splicing regulation that is not part of the basal
spliceosome is known to be undertaken by families of splicing regulatory proteins.
These splicing factors bind to signals in the vicinity of the exon and promote the
exon’s inclusion or exclusion by activating or inhibiting the function of the splice site.
The number of classes and characteristics of these regulatory proteins and their RNA
binding sites are relatively little known and are currently under active investigation
(Irimia and Roy, 2008; Pettigrew and Brown, 2008; Solis, et al., 2008).
24
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
These properties from alternative splicing should change the vision of gene regulatory
networks, defined by a fixed topology and a set of rules of interactions, by a more
dynamical concept.
Especially now, that it is accepted that alternative splicing is not the exception but the
rule for about 40-60% of the human genes (Downes, 2004; Stamm, 2002). However,
until now just a few information concerning alternative splicing regulation is available
and therefore it could not to be included into gene regulatory networks models. It
should be mentioned that alternative splicing is not a problem for the reverse
engineering of gene regulatory networks task by it self, rather for the prediction of
gene network behavior (see the difference on chapter 2 between inference power and
prediction power). It is, at the cellular level all signaling mechanisms are context
dependent, and what applies for a given cell at certain development stage, is not
applicable to another cell line or the same cell line at another cell cycle stage or under
different environmental conditions. Hence, one should carefully extrapolate what is
encountered in a cell line to another cell line or the same cell but under different
conditions. The distinction between these two situations will be further explained on
the 2nd Chapter.
1.5.2 RNA interference
Less than 2% of the human genome is translated into proteins (He, 2004), yet more
than 40% of the genome is thought to be transcribed into RNA (Ben-Dov, et al.,
2008). The vast fraction of untranslated RNA's includes several kinds of functional
non coding RNA, like snRNA (Spliceosomal and U7), snoRNA, telomerase RNA,
SRP RNA (protein trafficking), tRNA, TSK RNA (transcription elongation), and a
group of RNA that interfere with the expression of genes: siRNA, miRNA, piRNA.
This last family has different mechanisms to act, but share the characteristic of
silencing the expression of genes. The siRNA is oriented to silence exogenous genes
like those present in viruses (Juliano, et al., 2008) piRNA is utilized in mammalian
germ cells (Paddison, 2008) and finally the miRNA is utilized by plants and animals
cells to selectively silence genes during development (Boutros and Ahringer, 2008)
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
25
differentiation or proliferation as part of another level of gene regulation.
miRNA regulation is similar to gene expression, they have promoters and enhancers,
but also could be transcribed as part of some gene and later spliced during the mRNA
maturation. However, they are initially transcribed as pri-mRNA, removed by an
enzyme (drosha) in the nucleus, exported to the cytoplasm as pre-mRNA, where are
further removed by another enzyme (dicer) generating mature miRNA. It is there, in
the cytoplasm where miRNA´s finally meets its target mRNA, basically by WatsonCrick complementarities’. Once miRNAs meet their target mRNA, they decrease its
function by different possible ways like direct cleavage (mostly in plants), mRNA
deadenylation, affecting the stability of their target mRNA and inhibiting translation
(mainly in animals).
Currently there are 328 miRNA's annotated in the human genome (Chen and
Rajewsky, 2007), but is thought that there are more than 1000. Interestingly, even
though this miRNA is just 22 bases long they exhibit a relative highly conserved
sequence, and it has been shown that every class of miRNA could affect several
(hundreds) of different genes (He, 2004). It is thought that more than 30% of the
human genes are also regulated by miRNA’s. Therefore this mechanism should be
included explicit or implicitly by any model for gene regulatory networks.
Nevertheless, the problem to include this mechanism is that in animals the main
mechanism of gene repression through miRNA’s is in proteic translation repression. It
means, it does not affect the synthesis of mRNA and therefore would not be reflected
at the transcriptome level and in case, also would not be reflected in the mRNA
microchip technology.
Ideally, there should be information regarding the stability of mRNA, the relationship
of mRNA translated to proteins, the proteins half life, the state of activity of proteins
involved in gene regulation, localization of this proteins, complex formations etc. to
create a dynamical model of gene regulatory networks. But the case of miRNA
negative regulation is of particular relevance for reverse engineering of gene
regulatory networks, because it acts as a negative regulation of genes. Once again, the
distinction between both areas will be further explained in Chapter 2.
26
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
1.5.3 Dimensional in-homogeneities
As it has been mentioned before, the accessibility of the genes into the chromatin is a
general mechanism of gene expression repression (Cremer and Cremer, 2001).
However, chromatin structure is a field under investigation that proved that chromatin
has too many levels of control (Lanctôt, et al., 2007). A starting point to exemplify the
chromatin complexity is the super-coiling structure with its seven levels of packaging
that, according with the previously described, function as a basal repressive steric
barrier. Other gene expression regulatory mechanisms at the Chromatin level are the
chromosome territories, acethylation and methylation of histones, interaction with the
nuclear (actin) matrix, translocations of genes etc. In general all these processes
promote or inhibit gene activity by means of favoring or inhibiting diffusion of the
reactants (Misteli, 2001).
Additionally to the three dimensional in-homogeneities, there is another level of gene
regulation that is far beyond of the scope of the actual models. This additional
complexity comes from the fact that some of process previously described as
activation of TF, methylation and acethylation of histones are precisely regulated by
cytoplastmatic events called signal transduction. Cell signal transduction is a large
and very important area of study under intense research that is beyond the scope of
this work.
However, the emergent aspect to notice here is that the previously described inhomogenic diffusive processes, and the cell signal transduction processes of
communicates with each other through a fiscal barrier: the nuclear envelope
(Auboeuf, et al., 2007). This fiscal barrier controls with high precision the fluxes of
molecules between focalized processes and signal transduction by the means of
nuclear pores that often uses active transport (against concentration gradient, an
energy dependent process). Hence, a multi compartment modeling should be
applicable if simultaneous data of gene activity and signal transduction where
available. However, this also could be modeled by two means: reaction diffusion
models or Ordinary differential equations (ODE, from here on) with time delays.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
27
Reaction diffusion models are usually based on partial differential equations, (despite
some other possible approaches like cellular automata have being used) and have
shown to be very useful if the adequate data is available. However, to model the
multi-compartment fluxes coupled with diffusion reaction is, in general, difficult to be
modeled. On the other hand time delayed models are an alternative to model the final
effect of these complex processes as transport and focalizations of metabolites.
This good performance of time-delayed models has the price of not representing a
precise mechanics of the original system (kinetics constants) but instead they
represent the global behavior and structure through a semi-parametric model (have
being called phenomenological ones). However, depending on the goals of a research
project this last could be very useful.
28
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
29
2. Reverse engineering and modelling of genetic
network modules
2.1 Related work
The task of recovering the wiring between elements of a given system and the rules
governing their interaction, using data of its dynamical behavior is known as reverse
engineering. In the context of functional genomics, it means finding out which genes
regulate which others, how and under which circumstances. In other words, it is the
task of finding the topology of the gene regulatory network (GNR) related to a
particular cell line. Several experimental works and theoretical approaches have been
performed in this field. Since there are good reviews on the reverse engineering of
gene regulatory networks, in the present chapter I will just briefly describe some of
these works, starting with some common agreements in reverse engineering of gene
regulatory networks field, followed by a description of the most important theoretical
approaches on this field.
30
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
2.2 General concepts
Into a given cellular system, gene activity influences directly or indirectly the activity
of other genes. This system could be uni-cellular or pluri-cellular, where gene activity
from one cell influences the gene activity into another cell, as occurs in a tissue.
These interactions also could span the cell life cycle, relaying their influence to the
offspring cells. Moreover, this influence could be direct (by e.g. through miRNA), or
indirect upon the activity of the proteic sub-product of genes as TF or protein kinases.
However, more often the activity of one gene influences the activity of another as
means of the activity of a third gene. The reason for this behavior is that cellular
events occur sequentially, creating orchestrated cascades of gene activation as occurs
in differentiation processes. Additionally, activation of genes does not occur only
sequentially but also in parallel, meaning that often one gene influences more than
another one. Often some genes behave as hubs, where one gene influences several
target genes. The last part of this entangling process is the control of it, where the
activity of sub sequenced activated genes, in turn feedback influencing the activity of
the firstly activated genes. When this feedback process is positive, an amplification
occurs and is often used by the cellular system as a bi-stability control switch (Smolen
and Baxter, 2000). When the feedback is negative, a dampening occurs and is often
used by the cells for oscillatory or periodic behaviors as cell cycle or circadian
rhythms (Smolen and Baxter, 2000). In this way a network of precise regulations
emerges and is called a gene regulatory network (GRN).
In the systems biology community, the set of interactions (wirings and rules)
governing the behavior of a GRN is known as network topology. The task of finding
this network topology from the data of the dynamics of a given GRN is the goal of
reverse engineering. Usually, in order to perform the reverse engineering (RE, from
here on) task a formalism to represent the original system dynamics is needed, and is
called a model. The modeling of a given dynamic cellular system is very often not
clearly distinguished from the task of reverse engineering from that system. A model
is a representation of a system that helps us to understand its complexity.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
31
If the model is validated, by e.g. a cross validation technique, then it could be used to
make predictions about the original system. This last is known as the predictive
power (van Someren, et al., 2002) of a model. However, to model a system one needs
to know the wiring and rules of it topology. As we usually do not have complete
information about this topology and if there is information about the dynamic of that
system one can use the model to infer the missing topology. If the model is a good
representation of the original system, it has more probabilities to make the correct
inferences. This capability is known as the inference power (van Someren, et al.,
2002) of the model.
To perform the reverse engineering task based on dynamical data, series of gene
expression data over time (microarrays or RT-PCR) are used to represent a GRN
dynamics, as could be the entire genome of a cellular line. In this context every gene
from the GRN is known as a node (n) from a network, which is the GRN. The
measurement of how many elements are wired to a particular node is known as the
connectivity (K) of the node. The level of transcription for a given gene is known as
its activity state. Here the set of all (N) activity states measured at a given time point,
represents one transcriptional state of the system. In turn, a set of (system states)
measured time points of microarrays represents a trajectory of the cellular system.
Some works started by Kauffman (Iguchi, et al., 2007; Kauffman, 1969; Kauffman,
2004; Kauffman, et al., 2004; Kauffman, 1969; Socolar, 2003) on the late 60´s
proposed to see periodic behaviors as the cell cycle from a given cell line, as an
attractor of the entire organism’s genome. A phenomenological description of an
attractor could be the set of systems states where a system tends to exist. That is,
independently from the initial state a system have, it will tend to move into to the
closest attractor. This implies that, a system could have more than one attractor. The
vicinity from these attractors, where they applied their influence (like the size of a
funnel), is known as the basin of attractor. In this way, a multicelullar organism is the
entire system, and it different cell lines represent different attractors1 from that
organism.
1
There are different kinds of attractors like; the fixed-point attractor (the simplest)
limit cycles, toroids and strange attractors. Here, the analogy is between the cell cycle
and the limit cycles attractors.
32
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
When a system is observed evolving from an initial given state until it reaches an
attractor, it is said that the trajectory states are transition states. In this perspective,
the last goal of the RE of GRN works is, to achieve the necessary knowledge to
perturb a given cellular system to move it apart from its original attractor to another
desired one. For instance, the goal of RE could be to selectively control cells
belonging to a cancer attractor to differentiate them into a desired attractor, as could
be the apoptosis (programmed cell death).
2.3 Dimensionality reduction by data selection
When using microarray time series the system is highly undetermined. There is a
large lack of data in different dimensions, like observational time window, granularity
of measurements, diversity of conditions (stimulus response curves for different
stimuli and or conditions), repetitions, etc. Therefore, in order to reduce the
dimensionality of the system to be analyzed several works circumscribe the system to
be analyzed to the set of responding genes to a given stimuli. The criteria to select
those responding genes have evolved over the last years, and since it is a crucial step
for the reverse engineering of gene regulatory networks, some of them like those who
have been tested in this work, will be discussed below.
Threshold data selection from complete data sets
A common practice among biologist to reduce the gene data to be analyzed is to take
into account only those genes which expression has changed after a specific cellular
stimulus by at least two fold expressions, or by deleting genes with incomplete data or
those which standard deviation does not change beyond an arbitrary threshold.
Besides the choosing of a certain threshold it always will be arbitrary and this
approach faces two main drawbacks: (i) after correcting for multiple hypotheses
testing, no individual gene may meet the threshold for statistical significance, because
the relevant biological differences are modest relative to the noise inherent to the
microarray technology.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
33
(ii) Alternatively, one may be left with a long list of statistically significant genes
without any unifying biological theme. Interpretation can be daunting and ad hoc,
being dependent on a biologist’s area of expertise.
Functional modules
Due to the previous two problems on the threshold filtering of data, I will discuss here
different approaches oriented to reduce the dimensionality of the data. Three of them
are based on circumscribing the reverse engineering and modeling process to a
smaller functional module or cellular function. In this way the first main task to
perform those processes is to isolate the smaller number of state variables, able to
describe the given cellular module without missing important information. Therefore,
this module should be functionally self-sufficient and the available data reflect the
orthogonality (Lipan, 2005) from this module.
GSEA algorithm
The gene set enrichment (GSEA) algorithm is proposed (Subramanian, et al., 2005) to
be the solution of the complex task of isolating the important set of genes related to a
particular stimulus-response study. The logic of it is that single-gene analysis may
miss important effects on pathways, because cellular processes often affect sets of
genes acting in concert. They point that an increase of 20% in all genes encoding
members of a metabolic pathway may dramatically alter the flux through the pathway
and may be more important than a 20-fold increase in a single gene.
Gene Set Enrichment Analysis (GSEA) evaluates microarray data at the level of gene
sets. The gene sets are defined based on prior biological knowledge, e.g., published
information about biochemical pathways or co-expression in previous experiments.
The goal of GSEA is to determine whether members of a gene set tend to occur
toward the top (or bottom) of a list, in which case the gene set is correlated with the
phenotypic class distinction.
34
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Experimental isolation of modules through periodic stimulus
An experimental approach to the problem of isolation of functional genetic modules
has been proposed (Lipan, 2005) by using oscillatory inputs to analyze the response
of a cellular system.
Lipan and Wong propose in their work that an oscillatory input has many advantages:
(i) the measurements can be extended to encompass many periods so the signal-tonoise ratio can be dramatically improved; (ii) the measurement can start after transient
effects subside, so that the data become easier to incorporate into a coherent physical
model; and (iii) an oscillatory stimulus has more parameters (period, intensity, slopes
of the increasing and decreasing regimes of the stimulus) than a step stimulus. As a
consequence, the measured response will contain much more quantitative
information.
The genes that interact with the driven gene will be modulated by the input frequency.
The rest of the genes will have different expression profiles dictated by the internal
parameters of the biological system. This point of view is supported by their findings.
Lipan and Wong notice that the measured data can be expressed as a sum of
exponentially decaying functions, e-t, if a step stimulus is used while for a periodic
input the response contains only exponentials with imaginary argument, ei-t.
Mathematically, the main difference between exponentials with real arguments, e-t,
and those with imaginary arguments, ei-t, is that with the former one cannot form an
orthogonal basis of functions, whereas such a basis can be formed with the latter.
Therefore they propose that, in general, the response of the network to a step input
will be a sum of components that are not orthogonal on each other. The time
dependence of these non-orthogonal components can be more complex than an
exponential function; they can contain polynomials in time or decaying oscillations,
depending on the position in the complex plane of eigenvalues of a transfer matrix H.
In contrast, the permanent response obtained from a periodic input is a sum of Fourier
components that form an orthogonal set. Orthogonal components are much easier to
separate than non-orthogonal ones. This mathematical difference explains the
advantage of using oscillatory inputs.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
35
Clustering
Cluster analysis is an exploratory data analysis technique which aims at sorting
different objects into groups in a way that the degree of association between two
objects is maximal if they belong to the same group and minimal otherwise. Some
works like Wahde et al (Wahde, 2000), have follow the logical step to capture the
global behavior of the entire cellular system representing it by clusters instead of
representing particular genes, reducing in this way the dimensionality of the system to
be engineered.
However, in their work Wahde et al assumes that genes belonging to the same cluster
share the same function. Therefore, this approach proposes to use the centroid (mean
of the Euclidean distances between every gene expression profiles into a cluster) of
every cluster to represent the function related to that group of genes. In this way the
entire genome is represented by the biological functions that respond to the given
stimuli.
However, this approach faces two issues. The first is related to the cluster
interpretation. Genes belonging to a particular cluster could or not share the same
function. More likely, genes belonging to a given cluster share the same regulation,
but they could belong to different biological functions. The second issue of this
approach is that at the end, it is not possible to map the centroid to any particular
gene. Therefore the practical use of centroids to represent the entire genome is
limited.
36
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
2.4 Theoretical works
2.4.1 Boolean Networks
Almost 40 years ago Kauffman (1969) developed this model to represent genetic
networks. It is based on the assumption that genes exist basically in two possible
states; active (ON) or inactive (OFF). The activity state of each gene at a given time is
determined by a Boolean function (AND, OR, and NOT) of its inputs (wired nodes) at
the previous time step.
It is assumed that each gene is controlled by K other genes in the network. For this
models connectivity K is a very important parameter to determine the network
dynamics (with large K, the dynamics tends to be more chaotic). In Random Boolean
Network models, these K inputs, and a K-input Boolean function, are chosen at
random for each gene. The Boolean variables (ON/OFF states of the genes) at time
t+1 are determined by the state of the network at time t through the K inputs as well
as the logical function assigned to each gene. An excellent tool for calculating and
visualizing the dynamics of these networks is the DDLAB software (Wuensche,
1998). Under this approaches the total number of expression patterns is finite,
therefore the system will eventually return to an expression pattern that it has visited
earlier. Since the system is deterministic, it will keep following the exact same cycle
of expression patterns. This periodic state cycle is the previously defined attractor of
the network.
Another assumption performed by Kauffman, is that gene regulatory networks could
have a structure similar to random Boolean networks (RBN, from here on) and
therefore they should share some common features. For instance, the number of
distinct attractors of a Random Boolean Network tends to grow as a square root of the
number of nodes-genes (Kauffman, et al., 2004). If we equate the attractors of the
network with individual cell types, as Kauffman suggests, it is explained why a large
genome of a few billion base pairs is capable of a few hundred stable cell types.
This convergent behavior implies immense complexity reduction, convergence and
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
37
stabilization in networks of constrained architecture. In this way, with this model it is
possible to correlate the size and number of attractors from a RBN to the number of
cellular lines of an organism and the size of its genome in a predictive fashion. These
correlations are possible for a variety of species, according to some scaling
coefficients encountered by this work. Another insight into the behavior of large
regulatory networks is that given analogous Boolean functions (Kauffman, 1971) and
similar connectivity, RBN as well as biological systems tend to exhibit either a
maximum or a minimum of organization.
As recently the microarray technology comes to generate genome size data, this
model has been re-utilized to analyze dynamical properties of large GRN´s.
Some of these works utilized Boolean Networks (BN) to infer GRN from real data.
Liang and Somogyi (Liang, et al., 1998) have developed the REVEAL algorithm, that
utilized the Shannon entropy to correlate state transitions between nodes to infer the
mutual information between them, in this way and using a full search approach they
could the define the transitions rule table that reconstruct the original network
topology. The main drawbacks from this algorithm are the general criticism to BN of
representing gene expression by only two states: ON or OFF, and the assumed low
connectivity K.
Akutsu (Akutsu, et al., 2000) developed a series of different algorithms based on BN
to analyze the sample complexity of several variants networks, including noisy
Boolean networks. He proved that using a conceptual simpler approach, O(log2N)
random measurements are sufficient to identify a network of N genes with bounded
connectivity K. This means that for a data set with 1000 genes and a connectivity
K=2, in the order of only 10 independent measurements were sufficient for his
algorithm to infer the network topology. However, this approach has some strong
drawbacks. The first drawback is that its exhaustive search engine would utilize
O(1010) units of time. The second is that usually the connectivity is larger for real
GRN. And the last drawback is a general one for any Boolean network model: genes
expression exist in more than two ON/OF states, and very often the information
processing of this GRN are related to continuous levels of expression of their genes.
38
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
2.4.2 Differential equation systems
To overcome some of the limitations of the BN, some other works have been
developed to model the GRN on a continuous gene expression basis, using ordinary
differential equations (ODE). A set of ODEs, one for each gene, describes gene
regulation as a function of other genes:
dxi
dt = fi (x1 ,..., x N ,u,θ i )
2. 1
where xi(t) is the concentration of transcript i measured at time t, θi is a vector of
parameters describing interactions among genes (the edges of the graph), i =1 … N,
N is the number of genes and u is an external perturbation to the system.
As ODEs are deterministic, the interactions among genes represent causal
interactions, and not statistical dependencies as in other methods. To reverse-engineer
a network using ODEs means to choose a functional form for fi and then to estimate
the unknown parameters θi for each i from the gene expression data D using some
optimization technique.
With an ODE-based approach signed directed graphs are obtained and it can be
applied to both steady-state and time-series expression data. Another advantage of
using ODE approaches is that once the parameters θi, for all i are known, equation
(2.1) can be used to predict the behavior of the network under different conditions
(i.e. gene knockout, treatment with an external agent, etc.) as mentioned before as
prediction power.
There are many different approaches that can be enclosed into this differential
equation approaches, I will just briefly describe the major categories into which
particular models could be assigned and mention some examples: generalized additive
models, recurrent neural networks, S-systems, pair-wise equations.
Historically, systems of differential equations have long ago proved their validity in
modeling simple gene regulation systems. An example is the work of Mjolsness et al.
(1991), which used a hybrid homogeneous and 2D spatial reaction diffusion approach
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
39
to model a small number of genes involved in pattern formation during the blastoderm
stage of development in Drosophila (Reinitz, 1995). In this work, the change in
expression levels at each time point depended on a weighted sum of inputs from other
genes, and diffusion from neighboring “cells”. Synchronized cell divisions along a
longitudinal axis (under the control of a maternal clock) were alternated by updating
the gene expression levels. This model was able to successfully reproduce the pattern
of eve stripes in Drosophila, as well as some mutant patterns on which the model was
not explicitly trained.
After this model was introduced, several other models that use a similar formalism to
the so-called connectionist model (Mjolsness et al., 1991) were developed. Some
examples are the linear model (D'Haeseleer, et al., 1999), linear transcription model
(Chen, et al., 1999), weight matrix model (Weaver, et al., 1999) etc. Therefore, it has
being proposed (D´Haeseleer, 2000)to unify all of them by their common additive
nature and classified them into the so-called generalized additive models. Here I will
briefly describe some of them.
Linear system
In last term, it is possible to interpret these models as a multiple regression process:
dxi
w ji x j (t ) + bi
dt = ∑
j
2. 2
where xi is the level of expression of gene i at time t, bi is a bias term indicating the
basal expression of gene i when the summation of regulatory inputs is zero, and
weight wji indicates the strength of the influence of gene j on the regulation of gene i.
Therefore, given an equidistant time series of expression levels (or an equidistant
interpolation of a non-equidistant time series), it is possible to use linear algebra to
find the least-squares fit to the data.
Chen et al. (1999) presented a number of linear differential equation models, which
included both mRNA, and protein levels. They showed how such models can be
solved using linear algebra and Fourier transforms. Interestingly, they find that
40
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
mRNA concentrations alone are not sufficient to solve their model, without at least
the initial protein levels. Conversely, their model can be solved given only a time
series of protein concentrations.
D’Haeseleer (D'Haeseleer, et al., 2000) showed that even a simple linear model can
be used to infer biologically relevant regulatory relationships from real data sets. He
applied a linear model to the central nervous system differentiation rat data. It is one
of the few works developed over RT-PCR data, which are considerably more
quantitative that microarray data. In his work, he merged two data sets to obtain a new
data set with 65 genes and 28 time points. However, the dimensionality problem still
appears with several models able to fit equally well the data. To decrease this
dimensionality problem he used a spline interpolation scheme to obtain more data
points.
Van Someren et al.(van Someren, et al., 2000) combined a linear model with
clustering techniques to purpose a solution to the dimensionality problem. They
applied their model to the Yeast cell cycle data from Spellman.(Spellman, et al.,
1998) and showed that by working with the centroids of the clustered data, it was
possible to reconstruct a global network of the yeast cell cycle. However, as
previously mentioned, this approach has the drawback of not being able to correlate a
particular gene with a given cellular function.
Recurrent neural networks
As has been said, one of the pioneer’s works on the area using differential equations
is the one from Mjolsness et al (1991). However, this work as a series of other
developments could be seen as recurrent neural networks. Some advantages of seeing
these works under this perspective, is that the artificial intelligence community has
developed a series of good optimization methods for this kind of recurrent neural
networks, as back propagation through time and global optimization with genetic
algorithms (GA).
On the other hand, some works about continuous time recurrent neural networks
(CTRNN from here on) have proved that given enough data, there is just one model
that better fits the data. This uniqueness (Albertini and Sontag, 1993) property is
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
41
highly desired and in the RE of GRN has been called consistency.
Weaver et al. (1999) showed how a non-linear transfer function can be incorporated
into a linear model, and demonstrated that some randomly generated networks can be
accurately reconstructed using this modeling technique. To handle the dimensionality
problem, Weaver proposed the use of the Moore-Penrose pseudo-inverse.
This special matrix inverse produces a solution for undetermined problems that
minimizes the sum of the square weights but still perfectly fits the data. To impose a
limited connectivity, he proposed a greedy backward search that iteratively sets the
smallest weight to zero and then recomputed the pseudo-inverse on the, now slightly
less undetermined problem. However, this last technique is extremely sensitive to
noisy data.
Whade and Hertz (Wahde, 2000), inspired by the work from Mjolsness, utilized a
continuous time recurrent neural network model in the form:
dYi
−1
=
σ
W
Y
−
Y
+
θ
∑
ij
j
i
i
τ ,i = 1,...N,
dt j
to represent artificial as well as clustered GRN, where τ
2. 3
-1
are rate constants, Yi the
gene expression levels, dYi/dt the genes expression levels derivatives in time, θi the
genes basal expression levels, N the number of genes modeled, giving Wij as an NxN
(
weighting matrix and σ is the logistic sigmoid transfer function: σ (x ) = 1 + e− x
−1
)
.
Wahde and Hertz utilized a Genetic Algorithm as global optimization technique of
parameters and represent the same CNS rat data from Wen, by only four clusters.
Additionally, using artificial data they showed that it is better for the RE task, to have
multiple shorter time series than one long series.
Again, the major drawback of this work is that just four cellular functions without any
specific gene where taken into account.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
42
Recently, Hu (Hu, 2005) introduced a time delay term to the Wahde formulation in
order to account for additional process, as translation or diffusion that could delay the
response from a gene activity and its influence upon it target genes. They utilized the
Back propagation through time method to globally optimize their parameters. In they
work they reproduce the kinetics of six genes of the SOS DNA repair system of
E.coli. taking 50 time points data. In this way their model could recover seven of nine
experimentally reported regulations, as well as suggest six additional ones to be
tested. A drawback from this approach is the interpretation of the RNA decay, where
it is misunderstood and interpreted as the τ
-1
in a similar formalism as the equation
2.3.
S-system
S–systems (synergistic and saturable system) have long been used (Savageau, 1969)
as models of biochemical pathways, genetic networks and immune networks
(Akutsu, et al., 2000; Akutsu, et al., 2000). S–Systems are a class of non–linear
ordinary differential equations and have the form:
dXi
n
Xj
dt = α i ∏
j =1
gij
n
hij
− β∏ X j
2. 4
j =1
where n is the number of state variables or reactants Xi (X expressed in
concentration), and i,j (1 ≤ i,j ≤ n) are suffixes of state variables.
The terms gij and hij are the interactive effect of Xj to Xj. The first and second terms
represent all influences that increase and decrease Xi, respectively. The constants, gij,
and hij are exponential parameters referred to as kinetic orders. S–Systems have
unique mathematical properties allowing large realistic phenomena to be investigated
and can be derived from general mass balance equations by aggregating inputs and
outputs approximated by the products of power–law functions.
Each dimension of the S–System model represents the dynamics of a single variable
represented as the difference of two products of power–law functions, one describing
the influxes and the other describing the effluxes. The major disadvantage of the Ssystem is the large number of parameters to be estimated: 2X(X+1).
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
43
Pair-wise equations
Another way to overcome the so-called dimensionality problem is to restrict the
complexity of the model by only considering pair-wise relationships. Apparently,
Arkin (McAdams and Arkin, 1997) was the first to suggest the use of time-shifted
pair-wise correlations to model biochemical pathways. Initially, the position and
magnitude at which the maximal time-shifted cross-correlation occurs is computed in
a pair-wise fashion. Then a distance measure is constructed to perform hierarchical
clustering obtaining a linked tree of associated genes. Finally the model is completed
with information about directionality and time lags, and in turn it can provide
information about the dynamics of the system.
A further development comes with the work from Chen, who utilize a similar
approach but instead of using the correlation he performed the matching of peaks in
the expression profiles of genes. His algorithm performed threshold filtering followed
by the clustering step, obtaining profiles of sets of expression peaks. Then peaks in
the profiles are compared in a pair-wise fashion to determine causal activation scores.
From these scores a putative regulation network is constructed by optimizing it with a
simulating annealing approach.
Another related approach which combines the logical rules of Boolean network
models with some of the advantages of differential equation methods are “Glass
networks”. Glass networks have been proposed as a simplified model of genetic
networks (Edwards and Glass, 2000) as well as an underlying model for the reverse–
engineering of regulatory networks (Perkins, et al., 2006). The main drawback of
these models is the low connectivity K they are limited to, basically to single and in
some advanced cases to two pair interactions.
As a general criticism could be said that, differential equations presuppose that
concentrations of chemical species changes continuously and deterministically, both
of which assumptions may be questionable in the case of gene regulation (Gibson and
Mjolsness, 2001; Gillespie, 1977; McAdams and Arkin, 1999; Szallasi, 1999).
Against the first assumption is the fact that some of the components of GRN acts in
small numbers of molecules, as the transcription factors in the cell nucleus and a
44
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
single DNA molecule carrying the gene, compromising the continuity assumption.
Second, deterministic change presupposed by the use of the differential operator d=
dy/dt may be questionable due to fluctuations in the timing of cellular events, such as
the delay between start and finish of transcription. As a consequence, two regulatory
systems having the same initial conditions may enter into different states.
2.4.3 Stochastic Models
Stochastic models of gene regulatory processes claim to remedy many of the
drawbacks of deterministic (mainly differential equation) based approaches. One such
shortcoming is the assumption of a continuous rate of mRNA production. Typically,
transcription factors exist on very low concentrations in a cellular system, and this is
not well represented by continuous models as differential equations. In fact, as
exposed at the introduction, mRNA as well as proteins are not produced at a
continuous rate, but rather in short bursts (McAdams and Arkin, 1997). In addition,
some mechanisms of transcriptional regulation are known to amplify noise, creating
heterogeneity within a population. With the addition of noise in gene transcription,
individual cells may take different regulatory paths despite having the same
regulatory input (Guet, et al., 2002).
It is very likely that evolution has selected networks which can produce deterministic
behaviors from stochastic inputs in a noisy environment. In fact, certain topologies in
networks can attenuate the effects of noise (such as the mentioned control loops)
(Rao, et al., 2007) and also that noise can indeed act as a stabilizer itself in other
systems (Hasty, et al., 2000).
There are generally two methods for modeling stochastic gene regulation. The first
are stochastic differential equations:
dYi
dt = fi (Yi ) + ν i (t )
2. 5
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
45
This previous equation gives the form of a stochastic differential equation that
explicitly models noise in the system through the term ν (t). This equation is often
referred as the Langevin equation and in general is not analytical tractable. Typically,
solutions to the Langevin equations are obtained through the use of Monte–Carlo
algorithms. The conditions under which the approximation is valid may not always be
possible to satisfy in the case of genetic regulatory systems (de Jong, 2002).
The second approach is to characterize the transitions of a molecule using probability
functions. During each individual time step, a molecule is given a certain probability
of transitioning to a different discrete state. From this, a probability density function
for the behavior of the system can be obtained. Such systems are referred to as the
“Master Equation”. It has being proposed disregard the so-called master equation
altogether and directly simulate the time evolution of the regulatory system. This idea
underlies the stochastic simulation approach developed by Gillespie (Gillespie, 1977;
Gillespie, 1992).
Although stochastic models are often more realistic than their deterministic
counterparts, they are expensive to simulate. In fact, for many realistically sized
systems, stochastic approaches are impractical (Swain, et al., 2005). However,
stochastic models of gene regulation have been successfully used in Keasling
(Keasling, et al., 1995), Arkin (Arkin, et al., 1998) and Kastner (Kastner, et al., 2002)
just to mention a few examples. Recently, significant efforts have being performed to
reduce the computer simulation cost for the Gillespie algorithm (Cao and Gillespie,
2006; Slepoy, et al., 2008). However, their use for the RE of GRN area has not being
assessed.
2.4.4 Bayesian networks
Bayesian networks are probabilistic models. They model the conditional
independence structure between genes in the network. Edges in a Bayesian network
correspond to probabilistic dependence relations between nodes, described by
conditional probability distributions. Distributions used can be discrete or continuous,
and Bayesian networks can be used to compute likely successor states for a given
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
46
system in a known state.
A formal definition of Bayesian networks is:
A Bayesian Network is a directed, acyclic graph G = (X,A), together with a set of
local probability distributions P. The vertices X = {X1, ...,Xn} correspond to
variables, and the directed edges A represent probabilistic dependence relations
between the variables. If there is an arc from variable Xi to Xj , then Xj depends
probabilistically on Xi. In this case, Xi is called a parent of Xj. A node with no
parents is unconditional. P contains the local probability distributions of each node
Xi conditioned on its parents, p(Xi|parents(Xi) (Radde and Kaderali, 2007)).
In the formalism of Bayesian networks (Friedman, et al., 2000), the structure of a
genetic regulatory system is modeled by a directed acyclic graph G = (V; E). The
vertices i ∈V, 1 < i < n, represent genes or other elements and correspond to random
variables Xi. If i is a gene, then Xi will describe the expression level of i. For each Xi,
a conditional distribution p(Xi | parents(Xi)) is defined, where parents(Xi) denotes the
variables corresponding to the direct regulators of i in G.
In this approach the conditional independency i(Xi; Y | Z) express the fact that Xi is
independent of Y given Z, where Y and Z denote sets of variables. The graph encodes
the Markov assumption, stating that for every gene i in G, i(Xi; nondescendants (Xi ) |
parents.Xi )). By means of the Markov assumption, the joint probability distribution
can be decomposed into:
n
p (X ) = ∏ p (Xi | parents (Xi ))
2. 6
i =1
The resulting graphs from this Bayesian networks implies additional conditional
independencies. Two graphs, and hence two Bayesian networks, are said to be
equivalent, if they imply the same set of independencies. The graphs in an
equivalence class cannot be distinguished by observation on X. Equivalent graphs can
be formally characterized as having the same underlying undirected graph, but may
disagree on the direction of some of the edges see Friedman (Friedman, et al., 2000)
for details and references.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
47
Given a set of expression data D in the form of a set of independent values for X,
learning techniques for Bayesian networks allowed to some works to infer the
network, or rather the equivalence class of networks that best matches D.
These learning techniques rely on a matching score to evaluate the networks with
respect to the data and search for the network with the optimal score. As this
optimization problem is known to be NP-hard, heuristic search methods have to be
used, which are not guaranteed to lead to a globally optimal solution. However, an
additional problem is that currently available expression data underdetermines the
network, because just a few dozen of experiments provide information on the
transcription level of thousands of genes.
Friedman and colleagues (Friedman et al., 2000) proposed an heuristic algorithm for
the inference of Bayesian networks from expression data that is able to deal with this
so-called dimensionality problem. Instead of looking for a single network, or a single
equivalence class of networks, they focus on features that are common to high-scoring
networks. In particular, they look at Markov relations and order relations between
pairs of variables Xi and Xj. A Markov relation exists, if Xi is part of the minimal set
of variables that shields Xj from the rest of the variables, while an order relation
exists, if Xi is a parent of Xj in all of the graphs in an equivalence class. An order
relation between two variables may point at a causal relationship between the
corresponding genes. Statistical criteria to assess the confidence in the features have
been developed. A recent extension of the method (Pe'er, et al., 2001) is able to deal
with genetic mutations and considers additional features, like activation, inhibition,
and mediation relations between variables.
Markov relations and order relations have been studied in an application of the
algorithm to the cell cycle data set of Spellman and colleagues (Spellman, et al.,
1998) (see Pe’er et al. [2001] for another application). This data set contains 76
measurements of the mRNA expression level of 6,177 S. cerevisiae ORFs included in
time-series obtained under different cell cycle synchronization methods. The
Bayesian induction algorithm has been applied to the 800 genes whose expression
level varied over the cell cycle. By inspecting the high-confidence order relations in
48
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
the data, Friedman and colleagues found that only a few genes dominated the order,
which indicates that they are potential regulators of the cell cycle process.
Many of these genes are known to be involved in cell-cycle control and initiation. Of
the high-coné dense Markov relations, most pairs are functionally related. Some of
these relations were not revealed by the cluster analysis of Spellman and colleagues.
A Bayesian network approach towards modeling regulatory networks is attractive
because of its solid basis in statistics, which enables it to deal with the stochastic
aspects of gene expression and noisy measurements in a natural way. Moreover,
Bayesian networks can be used when only incomplete knowledge about the system is
available. Although Bayesian networks and the graph models are intuitive
representations of genetic regulatory networks, their disadvantage is to leave the
dynamical aspects of gene regulation implicit. To some extent, this can be overcome
through generalizations like dynamical Bayesian networks, which allow feedback
relations between genes to be modeled (Murphy, 1999).
Since this Dynamic Bayesian networks (DBN) are among of the more promising
works on the RE of GRN area, in this work I compare the performance of the here
introduced TDRNN model with the performance of DBN and the previously
explained CTRNN.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
49
50
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
3. Methods
3.1 Workflow
In few steps the working scheme could be described as: a) Times series data
acquisition. b) Data quality control and normalization. c) Data selection. To reduce
the number of state variables here I propose to focus on a small functional module or
cellular function. d) Data interpolation. This has been applied already in order to
reduce the solution space. e) Data fitting. The parameters of the model are globally
optimized by the use of a genetic algorithm (GA) to approximate the dynamics of the
selected module. f) Robust parameter identification. Statistical analysis is performed
to define the more likely parameters and consequently network connectivity g)
Summarization. The last step is the proposal of a network topology to describe the
dynamics of the original functional module. h) Error calculation. In case of the test
data, this error is calculated between the resultant network and the goal benchmark
network.
Figure 3.1 Workflow of the reverse engineering of gene regulatory networks. From left to right;
times series of data acquisition, normalization and interpolation of data, 50 regression
multiples to fit the data using a global parameter optimization approach, parameter
significance identification, and finally summarization of results on a network topology graph
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
51
3.2 Data pre-processing, Quality control
Given to the many sources of error of the microarray technology, a prerequisite for
working with experimental data from microarrays is to check their quality. This
includes checking for experimental outliers, discarding them, and checking for
statistical uncertainty of the results. In this thesis, quality control of every set of
microarray data has been performed with the package: simpleaffy (Wilson and Miller,
2005) from Bioconductor (Gentleman, et al., 2004) which is able to detect different
sources of errors during the different steps of the pipeline of the microarray data
production. For every set of times series, four different sources of errors are checked;
a) average background, b) 3-´5´ relationship and c) percentage of positive hybridizing
d) scale factor.
a) Average background: should be similar across all chips. There are several reasons
for a significant variation on the average background, but generally it is due to some
experimental problems, like having different concentrations of mRNA on the
hybridization cocktails, or a more efficient hybridization in some of the chips respect
to the others.
b) 3’—5’ relationship or early degradation of the mRNA: detected by an abnormal
signal from some control probe sets present on the chips. The chips contain some
specific genes probes, which have a particularly large sequence and well-defined
degradation kinetics. A change on this kinetic indicates that an abnormal degradation
of the sample has occurs and that measurements on the rest of the probe sets could be
lower than should be.
Most cell types ubiquitously express β-actin and GAPDH. These are relatively long
genes, and the majority of Affymetrix chips contain separate probesets targeting the
5’, mid and 3’ regions of their transcripts. By comparing the amount of signal from
the 3’ probeset to either the mid or 5’ probesets, it is possible to obtain a measure of
the quality of the RNA hybridized to the chip.
If the ratios are high then this indicates the presence of truncated transcripts. This may
occur if the in vitro transcription step has not performed well or if there is general
52
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
degradation of the RNA. Hence, the ratio of the 3’ and 5’ signal gives a measure of
RNA quality.
c) Percentage of positive hybridization: an indication of how many probes present on
the chip exhibit positive hybridizing. For a given experiment, one should expect that
the global number of positive hybridization is similar. However, this is not necessarily
the case because usually one is interested in finding the differences between
experiments. Therefore, problems with this evaluation should be taken carefully.
These percentages of responding genes (Present/Marginal/Absent calls) are generated
by looking at the difference between perfect matches PM and mismatches MM values
for each probe pair in a probeset. Probesets are flagged Marginal or absent when the
PM values for that probeset are not considered to be significantly above the MM
probes. As with scale factors, large differences between the numbers of genes called
present on different arrays can occur when varying amounts of labeled RNA have
been successfully hybridized to the chips. This can occur for similar reasons
(differences in array processing pipelines, variations in the amount of starting
material, etc.). The ‘% Present’ call simply represents the percentage of probesets
called Present on an array. As with Scale Factors, significant variations in % Present
call across the arrays in a study should be treated with caution. Note that usually the
absolute value is generally not a good metric because some cells naturally express
more genes than others
d) Scale factor: some normalization packages (e.g. MAS 5.0) adjust the mean value of
expression between different chips to the same value. If scale factors between arrays
are large, it is an indication of issues when trying to compare between chips.
The default normalization used by MAS 5.0 (and many other algorithms) makes the
assumption that gene expression does not change significantly for the vast majority of
transcripts in an experiment.
(Note that this assumption is also explicit in any analysis that looks for a relatively
small number of changing genes within a transcript population containing many
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
53
thousands (for example, looking for ~200 differentially expressed probesets from the
~54,000 found on the U133 plus 2 array).
One consequence of this is that the trimmed mean intensity for each array should be
constant, and by default, MAS 5.0 scales the intensity for every sample so that each
array has the same mean. The amount of scaling applied is represented by the ‘scale
factor’, which, therefore, provides a measure of the overall expression level for an
array, and (assuming all else remains constant), a reflection of how much labelled
RNA is hybridized to the chip. Large variations in scale factors signal cases where the
normalization assumptions are likely to fail due to issues with sample quality or
amount of starting material. Alternatively, they might occur if there have been
significant issues with RNA extraction, labeling, scanning or array manufacture. In
order to successfully compare data produced using different chips, Affymetrix
recommend that their scale factors should be within 3-fold of one another.
3.3 Data normalization
The goal of normalization is to be able to compare between different microarrays
chips. In general this is achieved by adjusting an average value of every experimental
array equal to that of the baseline array, in the case of time series data, the reference
array is that array (or those, in case of repetitions) without the external stimuli.
VSN data normalization
In the case of the microarrays chips (Affymetrix) used, their preprocessing involves
the following steps: a) combining the perfect match (PM) and mismatch (MM)
intensities in one number per probe, b) calibrating, c) transforming, and d)
summarizing the data. The algorithm vsn (Huber, et al., 2002), from the Bioconductor
platform, addresses the calibration and transformation steps.
This algorithm is considered to be the first choice for these tasks. The goal of this
algorithm is to provide robustness and avoid overfitting by first calibrating and then
performing a transformation of the data.
54
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Calibration is performed as follows:
Let yki be the matrix of uncalibrated data, with k indexing the rows and i the columns,
then the calibrated data y’ki is obtained by scaling with a factor λsi and shifting the
(Draper and Smith, 1998) data by the factor Osi,:
y 'ki =
yki − osi
3. 1
λsi
s is the so-called stratum (a classification of regions of the chips according to their
background signal) to probe k.
The transformation to a scale where the variance of the data is approximately
independent of the mean is performed by the use of the function:
h ki = arcsin h ( a0 + b0 y 'ki ) = log a0 + b0 y 'ki +
( a0 + b0 y 'ki )2 +1
3. 2
Where a and b are constants of proportionality calculated at the beginning of the
algorithm (here not shown). Both are applied simultaneously to the data in order to
obtain an almost constant transformed variance for every spot, independently from the
transformed mean for that spot. In this way, it is possible to work simultaneously with
high and low expression values, while at the same time avoiding any bias.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
55
3.4 Dimensionality problem. The use of interpolation
approaches
As it has been explained in the section 2.3 the system is highly undetermined due to
the lack of data in different dimensions, like time window, granularity, diversity of
conditions (stimulus response curves for different stimuli and or conditions),
repetitions, etc. Therefore, in order to increase the amount of data, here, it is assumed
that changes in gene expression between one measurement and the next one follow a
smooth function. Hence, interpolations in time are performed for every gene to obtain
a continuous set of data and to impose some constrains on the system.
Linear interpolation
Interpolation is a process for estimating values that lie between known data points.
The simplest way to perform interpolation consists in the use of a linear function to
produce continuous data points along the gap between two points using the shortest
trajectory. This kind of interpolation is named linear interpolation and supposes a
normal distribution of the measured data in relationship to the unknown real data.
If one have repetitions of experimental data measurements then it is possible to
estimate the dispersion of the data in the forms of variance or standard deviation of
gene expression. However, this variability would be the sum of two different
processes, the biological variability and the error involved in the measurement
process. The former fluctuations are usually hard to estimate, but the available
information points to a normal distribution of this kind of variability. In relationship
to the error associated to the experimental measurements, the supposition is that there
is not a systematic error associated, but in case, the Quality Control procedure
previously described should help to detect this problem.
56
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Cubic spline interpolation
Another possibility used in this work to increase the data is the cubic spline
interpolation. This algorithm also guarantees continuity of the interpolated data.
Therefore, the generated data could be differentiable everywhere and this
characteristic is often used by optimization methods like the gradient descendent in
other contexts. However, in any case interpolation with cubic splines implies the
smoothing of the original data.
There are two assumptions that underlie smoothing; a) the relationship between the
original data and the predicted data is smooth. b) The smoothing process results in a
smoothed value which is a better estimate of the original value because the noise has
been reduced. However, one should not fit data with a parametric model after
smoothing, because the act of smoothing invalidates the assumption that errors are
normally distributed (Draper and Smith, 1998). Nevertheless, I also tested and
compare the performance of this interpolation technique.
The Ziv Bar-Joseph et al. algorithm for gene expression representation
It occurs very often that experimental data is incomplete, very noisy and not
uniformly sampled. To overcome the missing data point problem, this algorithm (BarJoseph, 2004) splits the entire data set into clusters of genes showing similar behavior
and calculates the intra cluster noise. With this information, the algorithm calculates
the most likely values for the missing data. In this way, they provide with entire
vectors of data points of gene expression to a B-spline smoothing algorithm that also
takes into account the noise on the data obtaining a better representation of gene
expression behavior over time. This algorithm was tested while using experimental
data.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
57
3.5 Data fitting
As explained in the workflow section, the next step in reverse engineering basically
consists in performing multiple regressions to find the set of parameters of a given
model that best fits the available data. This last presupposes the existence of an
already chosen model to perform the regressions. Since the system is an ill posed
problem (Radde and Kaderali, 2007), due to the so-called dimensionality problem,
there are several models that could fit the data and one have to distinguish between
them (Often named system identification or inverse problem).
Here, two different groups of related models were tested: the CTRNN and the new
specific TDRNN here introduced. The data to be fitted is the data obtained from
interpolation of the time series gene expression as explained in the previous section.
The fitting is achieved mapping the behavior of every node from these models to a
specific gene from the original data set. Finally, since for every model there are
several sets of parameters that equally well fit the data, the next step is to discriminate
between those sets of parameters by performing statistical analysis to conclude with a
consistent solution to the reverse engineering problem; this will be covered in sections
3.6.3 and 4.14.
Error measuring, mean square error.
The fitting of the data by the models is measured by the use of the mean square error
between the output from every node and the data over time:
MSE = ∑ 0 ∑ i =0 ( d i − oi ) Li N
L
N
2
3. 3
Where (di - oi) is the error between the desired output and the obtained output from a
node at a particular time, N represents the total number of nodes studied and L the
simulation measured duration.
58
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Parameter optimization
This work uses a canonical genetic algorithm (Whitley, 1993) to globally optimize the
parameters of the models using a forward Euler2 integration scheme to simulate the
original system.
Genetic algorithms
A genetic algorithm (GA) is basically a search technique inspired in the nature of
evolution. It is often used to find an exact or approximate solution to optimization or
searching problems. Genetic algorithms are categorized as global optimization
heuristic searching technique.
A typical genetic algorithm requires two things to be defined:
1. a genetic representation of the solution domain (Blanco and Delgado, 2001),
2. a fitness function to evaluate the solution domain.
The fitness function is defined over the genetic representation and measures the
quality of the represented solution. The fitness function is always problem dependent.
Once I have defined the genetic representation and the fitness function, GA proceeds
to initialize a population of solutions randomly, then I applied repetitive adjustments
to the mutation, crossover, inversion and selection operators as described in table 3.1
until achieving a general increasing of the fitting.
Initially many individual solutions are randomly generated to form an initial
population. The population size depends on the nature of the problem, but typically
contains several hundreds or thousands of possible solutions. Traditionally, the
population is generated randomly, covering the entire range of possible solutions, the
so called search space. Occasionally, the solutions may be "seeded" in areas where
optimal solutions are likely to be found.
2
In general is a bad idea to perform serious calculations with Euler integration, here
the caution taken is the general recommendation in the computer science area to
choose an integration step at least ten times smaller than the smallest time constant.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
59
The next step consists of the evaluation of every individual from the initial population
according to the selected fitness function. Analogous to an evolution process, this
initial population is the first generation that will evolve through generations towards a
better solution to the problem in question. During each successive generation, a
proportion of the existing population is selected to breed a new generation. This one is
called the intermediary population. To generate the intermediary population,
individual solutions from the actual generation are selected through a fitness-based
process, where fitter solutions are preferentially selected. Usually the selection
methods (threshold ranked selection, elitist strategy) rank the fitness of each solution
and preferentially select the best solutions.
Other selection methods rate only a random sample of the population, this process
may be very time-consuming and also uses a fitness-based function. Most fitnessbased functions are stochastic and designed in a way so that a small proportion of less
fit solutions are selected. This helps keeping the diversity of the population large,
preventing premature convergence and therefore a poor final solution. Popular and
well-studied selection methods include roulette wheel selection and tournament
selection.
The next step is to generate a new-generation or population of solutions from those
selected by the use of the genetic operators: crossing-over (also called
recombination), and/or mutation.
Many crossing-over techniques exist for population’s individuals which use different
data structures to store themselves. In a single crossing over scheme, a crossover point
on both parents' organism strings is selected. All data beyond that point in either
organism string is swapped between the two parent organisms. The resulting
organisms are the children.
The classic example of a mutation operator involves a probability that an arbitrary bit
in a genetic sequence will be changed from its original state.
A common method of implementing the mutation operator involves generating a
random variable for each bit in a sequence. This random variable tells whether or not
a particular bit will be modified.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
60
For each new solution to be produced, a pair of "parents" solutions is selected from
the intermediary population to breed a new individual. By producing a "child"
solution using the above mentioned methods of crossover and mutation, a new
solution is created which typically shares many of the characteristics of its "parents".
New parents are selected for each child, and the process continues until a new
population of solutions of appropriate size is generated.
These processes ultimately result in the next generation population of chromosomes
which is different from the initial generation. Generally, the average fitness of every
new generation will increase since only the best organisms from the previous
generation are selected for breeding, along with a small proportion of less fit
solutions, for reasons already mentioned above.
This generational process is repeated until a termination condition has been reached.
Common terminating conditions are
•
A solution is found that satisfies minimum criteria
•
Fixed number of generations reached
•
The highest ranking solution's fitness is reaching or has reached a plateau such
that successive iterations no longer produce better results
•
Manual inspection
•
Combinations of the above.
Here binary encoding was used with four Bytes per parameter. The population size
was fixed at 1000 in all cases. Fitness rank-based selection was performed and the
stopping criterion was in all cases the number of generations (N=1000). Data was prescreened to conveniently adjust the mutation and crossing-over as shown in table 1 for
the different data sets.
Table 3.1 Optimization parameters adjustment
Repressilator
Initial Adjusted at gen:
Factor
Mut: 0.33
x 0.33
50, 150, 300, 750, 850
C.O. 0.4
+ 0.8
Mut = mutation rate
Yeast
Initial Adjusted at gen: Factor
0.5
x 0.2
100, 250, 500, 750
0.5
+ 0.9
C.O. = crossing over
gen = generation
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
61
Fitness rank-based selection was performed and the stopping criterion was in all cases
the number of generations.
Two optimization functions derived between the model outputs and the interpolated
data along the optimized period of time were used by the GA as fitness functions. The
first
−1
Fitness = (MSE + 1)
3. 4
optimization function was used to obtain a bounded fitness space, while the second
function:
Fitness = MSE −1 + λ e− β MSE
3. 5
additionally incorporates a fitness-adaptive weight pruning function (Bebis, 1996)
with β controlling the onset of the pruning starting point. λ is a function of the gene
interaction weights Wij according to the following parabolic function:
N2
λ=
∑ (aW
2
ij
i, j =1
+ bWij + c
) (dW + e),
ij
3. 6
Where N is the number of nodes and a, b, c, d and e (50, -170, 100, 5 and 1
respectively) are parameters controlling the shape of the parabolic function.
Figure 3. 2 Shape of the Lambda function that evaluates every weight to be
pruned. Weights between 1 and 3 are penalized while zero and values
higher than 3 are promoted
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
62
This function penalizes weights in the range [1,3] and promotes values close to 0 and
bigger than 3 to favor a sparse interaction matrix without affecting the MSE-derived
fitness. Changing the shape of the parabolic function or changing the range of
permissible interaction weights Wij did not change the results qualitatively.
Fifty optimization runs were performed for each of the synthetic and experimental
time series, randomly initializing the model parameters and the GA. The parameters
to be optimized were [τ ,ϑ ,θ,W ] and for the TDRNN additionally the time delay [δ ].
3.6 Models
3.6.1 The CTRNN model
In the continuous time recurrent neural network (CTRNN from here on) approach, it
is assumed that gene activity is reflected at the level of mRNA expression while
monitoring a cell stimulus-response or any other normal dynamical cellular processes.
The observables changes on mRNA expression are the balances of all those processes
described in the introduction, but projected at the mRNA expression level. The GRN
power inference of this model relays on the analogy between the continuous activity
level from it nodes and the continuously regulated gene expression of a given GRN.
For the internal structure of this model, - its nodes are predictors, from a multiple
regression point of view - it could be classified into the generalized linear models
(Bay, et al., 2002). This model could be derived from the gene activity analogy as
follows:
For a given gene (Yi), changes in mRNA expression over time are its synthesis (S)
and degradation (ϑ) balance:
dY
dt = S (Yi )− ϑ (Yi )
3. 7
Since very sparse information is available with respect to the mRNA degradation, this
model assumes a first order degradation kinetics for the gene expressions:
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
ϑ i (Yi ) = K i (Yi )
63
3. 8
In turn, it synthesis rate is the balance between direct or indirect interactions with
other genes (Yj) that activates or repress the gene in question:
Si (Yi ) = ∑ j =1 Y j
j=N
3. 9
An additional term, a constant bias (θj) is added to take into account a basal level of
activity for any gene:
dYi
dt = ∑ j =1 Y j − θ j − ϑ i (Yi )
j=N
(
)
3. 10
However, every interaction between two genes is a complex process following a nonlinear behavior. In gene regulatory networks this means a sensitive switch like
behavior as described by Hill’s kinetics (Hofmeyr, 1997; Setty, et al., 2003; Zaslaver,
et al., 2004). Therefore, on this family of generalized additive models, every node-tonode interaction is passed through a logistic sigmoid (σ) function:
σ ( x) =
1
1 + e− x
3. 11
giving us:
dYi
dt = ∑ j =1 σ Y j − θ j − ϑ i (Yi )
j=N
(
)
3. 12
The model here described represents the relative changes of gene expression rather
than being a model of mRNA concentration kinetics. The reason is that as explained
on the normalization section, the experimental data from the microarrays technology
is far from being quantitative mRNA´s concentrations. Microarray time series data are
rather strong indications of fold changes in relationship to a predefined state of
activity, usually respect to time cero defined as the time before any stimuli is supplied
to a cellular system.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
64
The application of the sigmoid function to every interaction and not to the entire
summation of them increases the precision of the calculation with respect to the
Whade and D´Haesseleer CTRNN models. At the same time, other consequence from
using the original Beer (Beer, 1995) CTRNN approach is, that a gene induction or
repression change scales proportional to the number of genes:
Since:
(
)
lim σ Y j − θ j = 1
(Y −θ )→∞
j
3. 13
j
and assuming that all interactions on the summation of a given node are activations
(positives) at their respective maximum value:
Yi (max)= lim ∑ σ (Y
N →∞
N
j =1
j
−θj = N
)
3. 14
b)
a)
1.0
Y2
0.8
5.0
5.0
2.5
2.5
0.0
0.0
0.0
-2.5
-2.5
0.6
σσ
-5.0
0.4
-5.0
Y1
0.5
1.0
1.5
2.0
σ
0.2
10
5
0
5
Y
Y
10
Figure 3.3 Sigmoid function comparison. Comparison among the output space after
applying the sigmoid function to the summation of all interacting terms in a) and the output
space after separated application of the sigmoid function to every interaction b). Here, for
simplicity in b) is depicted only to the second term of N.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
65
with N = number of total genes modeled by the CTRNN. Thus, the balance between
genes activating and inactivating each other acquires more relevance. This kind of
recurrent neural networks is full connected and explores every possible interaction
between nodes, in order to fit the data there are two possible tendencies for the final
topology: a) A mutual cancellation of interactions. b) the elimination of unnecessary
interactions giving a sparsely connected network. To the actual knowledge, this last is
the kind of network topologies present on the real genetic networks and has to be
favored by any model. Therefore, the second optimization function previously
described was used.
Additionally, a weighting (W) factor is applied to every interactions pair (ij), in order
to accentuate the importance of that interaction. Additionally, this weight assigns a
final sign to the interaction:
dYi N
= ∑ Wij σ Y j − θ j − ϑ iYi
dt
j =1
(
)
3.15
In order to integrate possible external information (for classification by e.g.) or
stimuli, an independent weighted term could be added:
dYi N
= ∑ Wij σ Y j − θ j − ϑ iYi + wI
dt
j =1
(
)
3.16
Finally, a general transcription rate parameter tau is added to describe the differences
on the response time from different classes of genes.
τ
dYi N
= ∑ Wijσ Y j − θ j − ϑ iYi +wI
dt
j =1
(
)
3.17
This model is exactly the CTRNN formalism known as continuous time recurrent
neural network (Beer). At the same time this is a series of coupled non-linear
differential equations. An advantage of seeing this model as CTRNN is that the work
from Funahashi et al (Funahashi, 1989) showed that they are universal aproximators
66
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
of smooth functions. Additionally, Albertini and Sontag (Albertini and Sontag, 1993)
showed that, if there is enough data, there is just one network topology that best fit the
data.
3.6.2 The TDRNN model
However, by the amount and quality of mRNA expression data required by the
CTRNN model there would never be enough. Another drawback of this model is that
it does not take into account that in Eukaryotes exist an important time delay from
gene activation until their product can interact with other genes. Here a modification
of eq. 3.17 was introduced by adding a delay (δi) to the interaction between genes
term:
τ
dYi N
= ∑ Wij σ Y j (t − δ i ) − θ j − ϑ iYi + wI
dt
j =1
((
) )
3.18
This time delayed recurrent neural network model (TDRNN) increases the nonlinearity with respect to the CTRNN, which non-linear parameter space behavior has
been analyzed by other works (Beer, 2006; Mathayomchan, 2002). More important is
that the differential time delay for every node moves this model apart from a Markov
chain process. Therefore no analytical solution is feasible. Hence this work will
follow a statistical approach to validate it in section 4.1.1. Moreover, despite this
abstraction is still far from integrating all gene regulation complexity, in this thesis, I
will demonstrate the utility of this novel model to represent gene regulatory networks
and in the RE task of them.
This TDRNN model is different from previously developed ones (Hu, et al., 2005;
Kim, 1998; Liao and Wang, 2003; Ma, 2004) in some senses. The model from Kim
demonstrates that TDRNN are superior than previous time delayed neural networks
(TDNN), adaptive time-delay neural network (ATNN) and multiple recurrent neural
networks (MRNN) for the tasks of temporal signal recognition, prediction and
identification. However, even though the general principle is similar, the delayed
information processing of his TDRNN is achieved trough the architecture of the
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
67
network, having input delay and output layers. Even though this architecture is
suitable for these purposes, it makes it impossible to do the mapping of one node to
one gene in a multiple regression model as the one here introduced. A similar
situation is present on the works from Ma and Liao and Wang, which are artificial
neural models developed for more general purposes. However, Ma demonstrates that
TDRNN are also capable to develop certain memory for spatio-temporal sequences.
On the other hand, this model is different from the model of Hu et al. in three
different aspects; a) This model explicitly model the decay rate associated to the
mRNA produced. b) The considered delays are constant for every gene, instead of
being particular for every interaction as proposed by Hu et al. One advantage of this is
having less parameters to estimate. However, more important is – as I consider - that
this is a more realistic situation, because the associated delay is due to translation and
diffusion of genes and proteins and it is constant among genes.
c) The non-linearity in my model is considered for every interaction instead as for the
entire summation. This marks the same difference as with previous CTRNN models.
In this work, I compared the performance of the TDRNN to infer GRN, respect to the
modified CTRNN version here exposed, the original utilized by Wahde and Hertz and
an available Dynamic Bayesian Network (DBN) implementation from Wu (Wu,
2004) in the GeneNetwork package.
3.6.3 Robust parameter determination
After performing the optimization runs, the weight matrix parameters were evaluated
from their z-score (D´Haeseleer et al., 2000),
zij =
W ij
σ ij
3.19
where W ij and σij denote the mean and the standard deviation of every matrix element
Wij is calculated from the 50 independent optimization runs. For the synthetic and
68
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
experimental GRN data robust parameters were defined as those having z-scores
greater than 2 and 1.5, respectively, corresponding to statistical significance values of
≈ 0.05 and ≈ 0.13375.
3.6.4 Graph generation and error distance measurements
A graph from the robust parameters determined on the previous step was generated,
by discretizing every robust parameter to a ternary representation according to values
between [1,-1] = 0; [5, 1] = 1; [-1,-5] = -1 generating in this way the, from here on,
so-called adjacency matrix. Then we used the Cytoscape (Shannon, et al., 2003)
facilities to generate a directed graph from every adjacency matrix.
With the adjacency matrix, the dissimilarity was calculated between it and the desired
benchmark network, the repressilator or the Yeast cell cycle, by the use of a directedweighted version of the graph edit distance algorithm (GED) developed by RoblesKelly et al. (Hancock, 2005; Robles-Kelly and Hancock, 2005). To calculate the
transformation cost between every resultant graph and the target network graph this
algorithm takes into account the existence of shortcuts (deletions-insertions) from the
semantic of directed weighted graphs. This is especially suitable for middle and large
size networks, and is the more objective way to compare between them.
3.6.5 Clustering of results
Cluster analysis simply discovers structures in data without explaining why they exist.
Therefore, cluster analysis methods are mostly used when we do not have any a priori
hypotheses, but they are still in the exploratory phase of research. Therefore,
statistical significance testing is really not appropriate here, even in cases when plevels are reported, as in k-means clustering. In a sense, cluster analysis finds the
"most significant solution possible."
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
69
In the RE area case, several dissimilar network topologies could fit the data equally
well. To distinguish between these solutions, clustering over the vector representation
of the matrix of weights from each experiment was performed. Using the Genesis
platform (Sturn, et al., 2002), a hierarchical and self organized maps clustering was
used to define the best partition number for a standardized k-means splitting
procedure. Then the splitting k-means algorithm was applied with the same
parameters for all the experiments to be compared. Robust parameters were identified
using the z-score method. To distinguish between these networks, the ratio between
the size and the mean fitness of every cluster was calculated.
An important prerequisite for clustering is the need of measuring the similarity or
dissimilarity between the elements of the sample to be grouped. There are several
different measurements of dissimilarity known as “distance” as Euclidean, Manhattan,
squared Euclidean, Chevichev, power distance etc. In this work the Euclidean
distance was utilized in all the cases where clustering is referred. It simply is the
geometric distance in the multidimensional space and it is computed as:
dissimilarity ( x, y ) = ∑ xi − yi
i
1
2
3.20
Being x and y the two different elements of the sample to be clustered where I want to
measure the dissimilarity and I represents the number of different dimensions
measured. Euclidean (and squared Euclidean) distances are usually computed from
raw data, and not from standardized data. This method has certain advantages (e.g.,
the distance between any two objects is not affected by the addition of new objects to
the analysis, which may be outliers). However, the distances can be greatly affected
by differences in scale among the dimensions from which the distances are computed.
Once the distances between elements of a given data sample were calculated, the
clustering algorithm groups or splits the sample into subgroups by mainly two
different strategies:
a) agglomerative techniques; starting from being every element on an isolated group,
those with the lowest distance are grouped together until ending with one unifying
cluster. This technique is typical for hierarchical clustering.
70
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
b) Splitting techniques; where given the initial set of data it is spliced according to it
elements similarity until ending with a desired number of clusters or a certain rule is
accomplished. This last technique is typical for k-means clustering.
Hierarchical Clustering
Hierarchical methods return a hierarchy of nested clusters, where each cluster
typically consists of the union of two or more smaller clusters. The hierarchical
methods can be further distinguished into agglomerative and divisive methods,
depending on whether they start with single object clusters and recursively merge
them into larger clusters, or start with the cluster containing all objects and
recursively divide it into smaller clusters.
K-means Partitioning
The k-means algorithm (MacQueen, 1967) can be used to partition N genes into K
clusters, where K is pre-determined by the user. Where K initial number of clusters is
chose by the user, and each distance among genes is calculated. Then starting from
the lowest K distances, every gene is assigned to the cluster with the nearest mean
named centroid. Next, the centroid for each cluster is recalculated as the average
expression pattern of all genes belonging to the cluster, and genes are reassigned to
the closest centroid. Membership in the clusters and cluster centroids are updated
iteratively until no more changes occur, or the amount of change falls below a predefined threshold. K-means clustering minimizes the sum of the squared distance to
the centroids, which tends to result in round clusters. Different random initial seeds
can be tried to assess the robustness of the clustering results.
Self –Organizing Maps clustering
The Self-Organized Map (SOM) method is closely related to k-means However, the
method is more structured than k-means in that way that the cluster centers are
located on a grid. In each iteration, a randomly selected gene expression pattern
attracts the nearest cluster center, plus some of its neighbors in the grid. Over time,
fewer cluster centers are updated at each iteration, until finally only the nearest cluster
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
71
is drawn towards each gene, placing the cluster centers in the center of gravity of the
surrounding expression patterns.
Drawbacks of this method are that the user has to specify a priori the number of
clusters (as for k-means), as well as the grid topology, including the dimensions of the
grid and the number of clusters in each dimension (e.g. 8 clusters could be mapped to
a 2x4 2D grid or a 2x2x2 3D cube). The artificial grid structure makes it very easy to
visualize the results, but may have residual effects on the final clustering.
3.6.6 Dynamic Bayesian Network
To compare the performance of the TDRNN with an established modeling
framework, the dynamic Bayesian network (DBN) approach implemented in the
GeneNetwork package (Wu et al., 2004) was used. The performance of the TDRNN
and DBN networks were compared on the same experimental data set (see section
4.2.1) under the same conditions: the networks were inferred from 100 linearly
interpolated data points. The GA implementation of the GeneNetwork package was
set to use a population size of 1000 individuals, running for 1000 generations with a
mutation rate and crossing over rate of 0.05 and 0.5, respectively. For the TDRNN the
optimization scheme as described in table 1 was used.
72
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
73
4. Results
This section is divided into three main parts. The first part contains the results
obtained with synthetic data of a system analogous to the so-called repressilator
synthetic system. Additionally, in this section, aspects are exposed to analyze the
introduced Time Delayed Recurrent Neural Network (TDRNN) model, through a
parallel study with a Continuous Time Recurrent Neural Network. This analysis and
the address uncovers some open questions in the reverse engineering area of gene
regulatory networks area as network sparsity, over-fitting and information required by
the here introduced model. In the second part the results obtained with the TDRNN
model of real data of the yeast (Saccharomyces cerevisiae) cell cycle are compared
with other approaches as the previously used CTRNN model and a Bayesian dynamic
network. In the third part, results related to the keratynocytes-fibroblast
communication system will be presented.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
74
4.1 Synthetic benchmark: The Repressilator
To assess the inference power of the TDRNN and CTRNN models, I tested both
models on generated synthetic data of the so-called repressilator system. The
repressilator system was chosen because it is among the simplest experimental
synthetic systems showing realistic characteristics of GRN as cyclic behavior. Cycles
occur often in biochemical networks and some of the more promising models for the
reverse engineering of GRN, the Bayesian networks (de Jong, 2002), cannot infer
cyclic networks. Furthermore, the repressilator work has become a good bench work
for different models on different areas (Elowitz, 2000).
The original repressilator is a synthetic GRN engineered in the E.coli bacteria, and
constitutes a network of three mutually repressing genes capable of undergoing limit
cycle oscillations. In this work, the repressilator dynamics is represented by three sine
waves derived from eq. 4.1 with a phase shift of 2 π between each of them.
3
(
)
y = sin x ⋅ 2π 150 ± 2 3 π ⋅ 5
4. 1
The sine curves have amplitude of 5 units and a period of 150 units to be in the period
time scale of the original work and on the expression amplitude scale of microarray
data as depicted in figure 4.1a.
Figure 4.1 Repressilator scheme. On a) is shown the oscillatory dynamics of the three
mutual repressing genes named repressilator b). This is a synthetically engineering system in
E.coli. The shadowed time elapsed in a) is assumed to be a relaxation lapse, therefore is not
included into the optimization process.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
75
With this synthetic data, the GA optimizations runs of the next sections were
performed. But in all cases, the first 200-simulation time units were discarded (see
figure 4.1a) to allow the model to reach an oscillatory steady state.
To represent the expression induction of every gene in the original experiment, the
recursive connection from every node is the only possibility in our working scheme.
Hence the goal network will consist of three nodes and six edges: auto-regulation of
each node and mutual repression in a cyclic way as depicted in figure 4.1 b. This has
the advantage of keeping the model from being driven by any arbitrary input function.
4.1.1 Parameter space selection
As has been stated, the biological systems to be engineered are highly undetermined,
and therefore exist an infinite number of parameter combinations able to equally well
fit the data. Additionally, the models tested here are semi parametric and in principle
have no bounds in the ranges their parameters could take. This models are focused to
perform the RE task at the network organization level rather than inferring kinetic
constants. Therefore, to choose the right parameter space is not a trivial task. In fact,
the chosen parameter is a compromise between different restrictions and objectives.
On the one hand, a large parameter space is desired in order to assure the convergence
between the model and the data. Additionally, some aspects of the parameter space
could be of particular relevance through a broader parameter range. In this case, it is
the range of the taus (τ) parameter, because the systems to be modeled have different
response time scales as the fast (in the order of minutes and probably seconds) signal
transduction process and the slower (in the order of hours) GRN. The range of the τ
parameter was chosen to be as broad as 3 ≤ τ ≤ 66 on the rest of this section. The
reason for this choice is that this result should also be valid in the next sections with
experimental data sets involving both kinds of processes.
Additionally, it is highly desirable to work under narrowed parameter space to speed
the searching task. But it is more important that some parameters should not correlate
with the RE task. Ideally, only the weights parameters should be correlated to a given
76
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
network topology. Therefore, it is highly desirable that the rest of the parameters exert
their influence mainly to fit the data, here, it is the bias parameter. In the absence of
any external stimuli, the individual biases are the only sources of dynamic behavior.
Actually, the effects of the biases and the external inputs on the locations of the
regions known as nullclines in the synaptic input space are identical (Beer 1995).
Therefore, considering these aspects, the chosen biases parameter space has the
−1 ≤ θ ≤ 1 short range.
Another parameter that should have a small effect on the RE task but plays an
important role to fit the data in a biologically inspired way, is the decay (υ)
parameter. However, since the present generalized additive models do not impose
upper and lower limits to the activity space of every node as it do the generalized
linear models, the maximum activity of every node scales with the network size
according to equation (3.18):
Yi (max)= lim ∑ σ (Y
N →∞
N
j =1
j
−θj = N
)
4. 2
Therefore, the υ decay parameter that balances the global activity of every node on
the
TDRNN
should
also
vary
proportionally.
Hence,
here
were
used
the Log10 (Yi ( max) ) , giving a 0 ≤ ϑ ≤ 0.5 range for the repressilator data set and
0 ≤ ϑ ≤ 3.5 for the experimental data set presented on the next sections.
Parameter screening
Finally, for the delay δ parameter, I performed a fast screening of the ranges this
parameter together with the taus τ parameter and beta pruning controlling factor could
take and I chose the range where the TDRNN model was performing better in respect
to the MSE and to the RE task. To evaluate this, the resultant weights of every
optimization run of the screening was discretized to a ternary representation: [1,-1] =
0; [5,1] = 1; [-1,-5] = -1 and compared with the goal network adjacent matrix of figure
3.1b. The screening was performed with a population size of 100 individuals. The
results of this screening are scatter-plotted on figure 4.2.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Beta (log)
Fitness
400
200
Delay (intervals)
400
2
10
0
0
4
10
-1
10
0
2
10
4
10
c)
5
10 0
10
2
4
10
g)
0
10
0
10
0
20
f)
5
10
10
-1
0
0
5
10
10
10
0
g)
e)
10
10
0
0
20
-1
10
Taus (intervals)
400
200
b)
Errors
d)
200
0 0
10
MSE (log)
a)
77
10
0
0
5
10
i)
0
5
10
20
10
0
5
10
Figure 4.2 Scatter-plot of the parameter space screening. On a, b and c panels are the
scatterplots of the influence of pruning over the fitness and inference power of the model. In
the middle vertical panels d, e and f, is represented the influence of the delay parameter of
the TDRNN model over the same performance indicators fitness and errors. Panels g, h and i
shown analogous information for the influence of the taus τ parameter. In dashed lines are
represented the asymptotic or oscillatory tendency that some parameters to follows.
On figure 4.2 a, b, and c is possible to see the functioning of the pruning function. On
4.2.a values of the fitness function clearly decrease for larger values of the Beta β
pruning controlling factor of equation 3.5. Notice that for β =102, β plays almost no
role and for β =103 does not play any role at all. On the panels 4.2 b and c is shown
that the MSE and the inference power (expressed in number of errors) are not affected
by the pruning process (β ≤ 102) despite the fitness varies with the beta factor. On
4.2.b, it is possible to see two distributions; on the upper part are runs with a good fit
(MSE on logarithmic scale) and on the lower part are runs with a poor fit of the data.
However, this bimodal distribution is present for all Beta values. On 4.2.c, the scatterplot suggests that those runs with no pruning (β ≥ 102) are only slightly more sensitive
concerning the inference power, showing no runs with one error.
On the scatterplots of the delay (δ) parameter ranges (figures 4.2. e and f), it is
possible to observe the same bimodal distribution previously observed. Here, it is
78
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
possible to see that the inference power of the model for a delay range of six units is
in a maximum performance zone. This apparently oscillatory behavior of the
parameters ranges is clearly related to the high symmetry of the repressilator system.
As demonstrated by the works of Beer (1995, 2006), the parameter space of the
CTRNN for such a dynamical system is divided into regions of topologicallyequivalent dynamics by bifurcation manifolds. Hence, for the reasons previously
exposed, it was chosen to work with the shortest delay interval of six units showing
the best performance.
An analogous situation to the delay parameter could be observed on the figures 4.2 g,
h and i for the taus ( τ ) parameter. However, for the reasons previously exposed, here
it was chosen to work with a broader parameter space to observe the implications of
this cyclic behavior in just one of the parameters. Ideally this parameter should work
on a broader space to cover diverse biological systems. However, since the others
parameters are constrained to a smaller space, the manifolds should not appear. Under
these circumstances, problems related to the solution of the systems of stiff
differential equations could rise, see next sections.
Parameter correlations and inference power
To corroborate that the inference power of the TDRNN and CTRNN models is mostly
insensitive with respect to the chosen parameters ranges ( 3 ≤ τ ≤ 66, −1 ≤ θ ≤ 1 ,
−1 ≤ υ ≤ 1 , −5 ≤ W ≤ 5 ), I calculated all correlations in between parameters, to the
MSE and the inference power (as means of errors). For this purpose, 150 optimization
runs under the previous standardized conditions were performed with the TDRNN
model using a 0 ≤ δ ≤ 6 range. I used a population of 100 individual’s size. To
discard spurious correlations due to none fitting problems (as observed by the
bimodal MSE distributions) a MSE= 0.4 was used as stopping criteria (see figure
4.3a). This is a normal optimization procedure when most of the runs share a similar
MSE.
To avoid interference, the pruning function was not used. On figure 4.3 the global
distributions of the parameters are depicted by their histograms (figure 4.3 c, d, e, f
and g).
M
or
e
10
4.
02
-25
Lambda
30
25
0
Errors
50
3.
21
8
6
4
2
0
70
75
2.
41
20
110
90
125
1.
61
40
79
c)
0.
80
Frequency
60
0
Fitness (MSE)
175
b)
80
10
a)
Frequency
200
150
100
50
0
0.
03
1.
68
3.
33
Frequency
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
-10
140
120
100
80
60
40
20
0
TAUS
f)
Frequency
50
DELAY
40
30
20
10
60
50
g)
40
30
20
10
0
-0
.9
9
-0
.8
0
-0
.6
1
-0
.4
2
-0
.2
3
-0
.0
4
0.
15
0.
34
0.
53
0.
72
0.
90
0
BIAS
0.
02
0.
11
0.
19
0.
28
0.
36
0.
45
0.
53
0.
62
0.
70
0.
79
0.
87
Frequency
60
e)
1.
00
1.
48
1.
95
2.
43
2.
90
3.
38
3.
86
4.
33
4.
81
5.
29
5.
76
Frequency
d)
2.
87
5.
05
7.
23
9.
41
11
.6
0
13
.7
8
15
.9
6
18
.1
4
20
.3
2
22
.5
0
24
.6
8
Frequency
| WEIGHTS |
80
70
60
50
40
30
20
10
0
DECAY
Figure 4. 3 Parameters histograms
The distributions depicted on figures 4.3 e, f and g, suggest a normal distribution for
the global decay, delay and bias parameters; this was corroborated by a KolmogorovSmirnov test of normality (p ≤ 0.05). Interestingly, figure 4.3.c shows a bimodal
distribution of the absolute value of weights. One of the processes appears to be close
to zero while the other is centered on interactions around 3 units. These are the
expected distributions for a network using 2/3 of it full connectivity as it occurs in the
repressilator topology. Superposed, in figure 4.3.c are the values of the Lambda
function (black bars using the second y axis scale at the right side) for the same
weight values. Notice that the shape of the λ function, exactly penalize the weights
between the two distributions, at the same time this λ function promote the moving of
the two distributions towards zero and larger than 3 units values respectively.
This pruning distribution is desired because it is easy to probe correlations with large
interactions, and obviously no interaction is the most desired feature of the pruning
function.
80
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Table 4.1 Pearson correlation among parameters and fitness (MSE) and inference power
(ERS).
ERS = number of errors
T=taus
B = bias
C= decay
D=delay
On the 4.3.d figure, it is clearly shown that the τ parameter have a Log normal like3
distribution skewed to the right side, with a median toward short values (the peak is
around 7 units). This shows that the model does not have any problems integrating
stiff equations systems; instead, it automatically chose those values that promoted its
stability according to the period size on the approximated sine functions.
Since the size of the sample was bigger than 100 elements, deviations from normality
on the previous distributions are less important (Hill and Lewicki, 2006), therefore
Pearson correlation (p-value ≤ 0.05) was calculated between every parameter pair as
well as to the MSE and the inference power (expressed in errors). The resultant
correlations are in table 4.1
3
The lognormal distribution could be used to model the time required to perform
some task when "large" values sometimes occur. It is always skewed to the right and
it has a longer right tail than the gamma or Weibull distributions. The lognormal
distribution is closely related to the classical normal distribution
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
81
On the 2nd and 3rd columns of the table 4.1, are shadowed the two more important
correlation series. The first is the correlation between the nodes individual parameters
and the fitness (MSE 3rd column) and the second between the same nodes individual
parameters and the inference power represented by the individual-errors (ERS, 2nd
column header). Here, it is demonstrated that no other parameter than the weights
could exert a stronger influence to the inference power of the model. Particularly, at
the ERS column, it is quite obvious that the more important weights concerning the
inference power are those which should be eliminated (W13,W21 and W32) to obtain
the repressilator topology.
By contrast, the bias (B1 and B3 rows on the table 4.1) is clearly the parameter that
correlates more strongly to the fitness (MSE column) of the model. These results
corroborate the assumption that only the weights correlate to the inference power
while the rest of the parameters mostly correlate to the fitting of the data. This result
is of high importance because of results from sections 4.1, 4.2 and 4.3; the analyses
will be focused on the square matrices of the weights (Wij).
One can argue that the weights should correlate strongly to the inference power
4
since it is calculated exactly on the weights. However, this is not true since the errors
are calculated in relationship to the ternary discretization of the weights. This
decreases the variance of the errors respect to the weights; therefore, this is not an
auto correlation measurement. Moreover, it is important to notice that even some
weights almost do not correlate to the errors, as W13, W21 and W32 - which are exactly
those of the mutual inhibitions - despite they are strongly related to the desired
topology.
Notice that in general there are mostly low correlations between the parameters and
the fitness and the inference power as one can see at the first two columns of the
correlations in the 4.1 table. Again, one can argue that it is that way because the
distribution of the fitness is monotonic with a low variance, as the MSE was prefixed
as stopping criteria. Nevertheless, this is not the case of the inference power. On
figure 4.3.b, the histogram of the errors distribution is shown and it clearly shows that
4
the inference power here refers to the inverse of errors
82
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
errors have a geometric distribution5 as opposed to the MSE. This means that actually,
one should expect stronger correlations for the inference power to any other column
than the expected for the fitness, but only the weights have this characteristic.
Instead of strong correlations between parameters and fitness or inference power, on
table 4.1 there are strong correlations (shown with grey scale background) between
the parameters and the weights (depicted on the 3rd quadrant of table 4.1).
Particularly, the strongest interactions are between the decay parameter and all the
edges of a given node (depicted in respective node’s boxes in table 4.1). In decreasing
order of correlation strength, those with the stronger correlations, like r ≥ 0.9, (cells
with black background and white numbers) are the direct correlations between the
positive auto-feedback loop of every node (W11,W22 and W33, see the weights scheme
into the table 4.1) and their respective decay parameter (see Figure 4.4 a, b and c).
These correlations are expected for this model in order to stabilize its outputs. So, for
a bigger positive auto-feedback loop a bigger stabilization decay parameter is needed.
5
The geometric distribution (discrete) with probability = p can be thought of as the
distribution of the number of failures before the first success in a sequence of
independent Bernoulli trials, where success occurs on each trial with a probability of
p and failure occurs on each trial with a probability of 1 - p.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
5
a)
W11
5
4
4
3
3
2
2
1
1
0
0
6
b)
W22
83
c)
W33
5
4
-1
3
2
1
0
0
0.5
0
0.5
DECAY 1
1
DECAY 1
-1
d)
DECAY 2
0
-1
-2
-3
-4
-5
-3
-4
-5
W12 (-)
g)
W13
1
0.5
DECAY 3
1
f)
W31 (-)
W23 (-)
W21
4
h)
4
i)
W32
3
3
2
2
2
1
1
1
0
-2
DECAY 3
-6
3
-1 0
0
1
-2
-4
0.5
-2
e)
-1
-3
4
0.5
DECAY 2 1
0
-2
-5
0.5
0
1
0
-1
0
-1 0
0.5
1
DECAY 1
0
-1 0
0
0
-1
0.5
1
DECAY 2
-2
0.5
1
DECAY 3
Figure 4. 4 Principal interactions scatterplots
The next strength correlation level (0.5 ≤r ≤0.6) for interactions between the decay
parameter and weights corresponds to those weights of the mutual repression (W12,
W23 and W31) edges. Particularly, the correlations are between the decay parameter of
a given node and the incoming repression edge from another node. Notice that since
those weights are negative, the correlation is a direct and not an anti-correlation as
one should expect for two processes that act together decreasing the node activity.
The last mentioned process is clarified on the scatterplots of figures 4.4 d, e and f,
where the angle of the correlation trend is negative, shown in the figure with dashed
lines.
Since no important correlation conclusions should be done based just on Pearson
correlations, a scatter-plot visual examination has been performed for all the
interaction pairs between parameters to check for any possible omission or spurious
correlation, because a nonlinear correlation could occur, or spurious correlations due
to bimodal distributions also could takes place and both cases could been mishandled
by a linear (Pearson) correlation.
84
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Additionally to the mentioned upper panels scatterplots of figure 4.4, the rest of the
more important scatter plots are shown in figures 4.4 middle panels (d, e and f) and
4.4 lower panels (g, h and i). In figure 4.4 middle panels, just linear correlations
(dashed lines) or anti-correlations were found for all cases, showing that even the time
delay increases the nonlinear behavior of every node’s activity, the global behavior is
far from being unpredictable.
Finally, in respect to the decay parameter and with an anti-correlation (-0.5≤ r ≤ -0.4),
lie those correlations between the decay parameter of a given node and the weight of
the incoming edge from another node. Notice that there are just two incoming edges
per node: the repressive one and the other one, this last is mostly working as an
activator. In this case, the reference is to the W13, W21 and W32 weights that should
not exist or being slightly positive.
Proceeding with an analogous study for another parameter, I found on table 4.1 that
the highest level of correlation strength (0.6 ≤ r ≤ 0.9) corresponds to the correlation
between the taus (τ) parameters of a given node and the incoming activation from
another node to the first one (see the figure on first quadrant of table 4.1.) Notice that
this is the edge that should not exist. According to this, the stronger the incoming
activator weight the larger the τ parameter, meaning a smaller expression ratio of the
node in question. The same situation occurs for the incoming negative repression
edges (W12, W23 and W31) and the τ parameter of every node, but at lower correlation
strength (0.2 ≤ r ≤ 0.7). These two correlations are logically needed to stabilize the socalled node’s reactivity or reaction rate. The larger the weights strength is, the larger
is the τ parameter which needs to be on the equation 3.18 to stabilize the global
activity of the node:
dYi 1 N
= ∑ Wijσ (Y j ( t − δ i ) ) − θ j − ϑiYi + wI
dt τ j =1
(
)
4.3
However, for the interactions between the τ and the auto regulatory edge weights
(W11, W22 and W33), the weak range (-0.5 ≤ r ≤ -0.3) of anti-correlation is partially
explained by the fact that, in order to produce an oscillatory behavior, this kind of
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
85
dynamical systems requires amplification. In this sense, the shorter the taus parameter
is, it needs a larger auto regulatory edge weight to keep the system oscillating.
Notice that the last two parameters analyzed are those which are not passed trough the
sigmoid function. This it is very logical because their range of influence over the
entire node activity is in the same order than the weights. Therefore a similar strong
influence is expected from the external I input term from the previous equation 4.3 (or
3.18). From here on, one has to be very cautious while analyzing this model with an
external input because then it could be not possible to separate its influence on the
Matrix of weights (Wij) as it is proved in this study.
Without pretending to create an unnecessary statistical model about the TDRNN
model topology, the next analysis was looking at the scatter plots of every pair of
interaction among weights in order to corroborate the absence of non-linear
interactions among them. The relevant correlation results are plotted in figure 4.5,
where the encountered strongest correlations on the 4th quadrant of table 4.1 were
those without a normal distribution of the weights.
Notice that all but two correlations (W32-W33 and W31-W35) show a linear correlation
(dashed trend lines) among them. However, the two correlations that are better
represented by a second order trend line are those with sign transitions in the weights,
while those showing all the data distributed in one quadrant show a linear correlation.
W12 5
3
1
-5
-1 0
W11 5 -5
-3
-5
-5
W13 5
4
3
2
1
0
-1 0
-2
-3
-4
-5
W11 5 -5
W23 5
4
3
2
1
0
-1 0
-2
-3
-4
-5
W22 5
4
3
2
1
0
-1 0
-2
-3
-4
-5
W22 5 -5
W21 5 -5
Figure 4. 5 Scatter-plot of the principal weights interactions
W31 5
4
3
2
1
0
-1 0
-2
-3
-4
-5
W32 5
4
3
2
1
0
-1 0
-2
-3
-4
-5
W33 5
5
W33
86
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
All this results are consistent with the expected behavior according to the works from
Beer (1995, 2006). More important is the fact that this results corroborate that the
inference power of the TDRNN and CTRNN models is mostly insensitive concerning
the chosen parameters ranges ( 3 ≤ τ ≤ 66, −1 ≤ θ ≤ 1 , −1 ≤ υ ≤ 1 , −5 ≤ W ≤ 5 ) of the
former results, restricting in this way the further analysis of the Wij matrices of
weights as described in the methods section.
4.1.2 Required data length.
To determine the data required to reverse engineer the cyclic repressilator system, the
model was simulated for 0.2, 0.33, 0.5, 0.6, 0.75, 1.0, 1.5, 2, and 3 oscillation periods
after an initial transient of T = 200 units. The 50 optimization runs were filtered from
outliers (MSE-mean +/- 2 standard deviations) and the robust parameters and
adjacency matrices were identified for each of the 9 simulation intervals as described
in methods. To compare the performance of the networks, the errors were defined as
the numbers of mismatches between the adjacency matrices of the inferred and the
goal network showed in figure 3.1b.
The results are plotted on Fig. 4.6 as a comparison between the errors of the TDRNN
and the CTRNN models for different simulation runs using weight pruning (P) or
unbiased parameter selection (not pruning - NP). Summing over all errors from all
simulations of nine time intervals the TDRNN-NP model performed best, having only
three errors. In comparison, the reverse engineering using the CTRNN-P, CTRNN-NP
and TDRNN-P gave a total of 11, 7 and 10 errors, respectively.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
87
Errors 6
4
2
0
0.2 0.33 0.5 0.6 0.75
CTRNN-P
TDRNN-P
1
1.5
2
CTRNN-NP
3
Periods
TDRNN-NP
Figure 4. 6 Inference power comparison along nine different optimized period intervals
between the TDRNN and CTRNN models with and without parameter pruning (P, NP). The
comparison is expressed in terms of errors, defined as mismatches between the robust
parameters calculated by any model and the topology of the goal network. For 0.5 periods, all
models can infer the topology without errors, but below and higher than 1.5 periods, NP and
TDRNN perform best.
In order to interpret the previous result in the light of the data fitting by every model,
a comparison of performance to fit the data by the four different model-reverse
engineering combinations (in the following called models for simplicity) is shown in
the boxplots of figure 4.7 This boxplots produces a box and a whisker plot for each of
the simulated periods of every model. The box has lines at the lower quartile, median
and upper quartile values. The whiskers extend from each end of the box to the
adjacent values in the data of the fittings, the most extreme values within 1.5 times,
the interquartile range from the ends of the box. In this boxplots, outliers are those
fittings with values beyond the ends of the whiskers. Outliers are displayed with a +
sign. Notches display the variability of the median between samples. The width of a
notch is computed so that boxplots whose notches do not overlap have different
medians at the 5% significance level. The significance level is based on a normal
distribution assumption, but comparisons of medians are reasonably robust for other
distributions. Comparing boxplot medians is like a visual hypothesis test, analogous
to the t test used for means.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
88
CTRNN-NP
CTRNN-P
0
MSE(log)
MSE (log)
0
10
-2
-2
10
10
0.2 0.3 0.5 0.6 0.75 1 1.5 2
Periods
TDRNN-NP
3
0.2 0.3 0.5 0.6 0.75 1 1.5 2
Periods
TDRNN-P
3
0.2 0.3 0.5 0.6 0.75 1 1.5
Periods
3
0
0
10
10
MSE(log)
MSE (log)
10
-2
-2
10
10
0.2 0.3 0.5 0.6 0.75 1 1.5 2
Periods
3
2
Figure 4.7 Data fits comparison among the TDRNN and CTRNN along different optimized
period intervals using (P) and not (NP) the adaptive pruning function. For every group an
easy-to-fit zone is identified into the dashed line box. Outliers of every group distribution are
represented by “+” sing.
The results showed in figure 4.7 clearly show 3 zones (separated by vertical dashed
lines) of fitness values for all the models. In all cases, the central region comprising
the 0.5, 0.6 and 0.75 period intervals, are the regions with better median fitting of the
data. These regions are called henceforth easy to fit regions. However, notice that in
all cases the inferior whiskers (extreme data) follow a clear decrement with the
quantity (expressed in period fractions) of fitted data. In this way, for all models the
extreme lower fitting is close to 0. This last result is logical and was expected, by
contrast, the easy to fit regions were an unexpected result.
Additionally to the boxplots a normality distribution test was performed (Lillie test, p
< 0.05), and since the distribution of the MSE deviated from normality between all
models, a parameter free test (Kruskal-Wallis test, p-value < 0.05) was performed to
compare between MSEs. Here I found that the TDRNN-NP model fitted the data
significantly better in 7 out of the 9 intervals. In figure 4.8 are shown the respective
boxplots showing the same result favourable for the TDRNN model.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
89
0
10
MSE (log)
10
0
10
CTnp CTp TDnp TDp
-2
CTnp CTp TDnp TDp
CTnp CTp TDnp TDp
MSE
0.08
0.03
0.03
0.02
0.02
0.04
0.01
0.01
0.02
0
CTnp CTp TDnp TDp
0
0.06
0
CTnp CTp TDnp TDp
CTnp CTp TDnp TDp
0.06
0.06
MSE
0.04
0.04
0.02
0.02
0
0
CTnp CTp TDnp TDp
0.04
0.02
CTnp CTp TDnp TDp
0
CTnp CTp TDnp TDp
Figure 4.8 Comparison of the fitting by the two models using (p) and not (np) the pruning
function at every optimized time window. CT = Continuous time recurrent neural network, TD
= Time delayed recurrent neural network.
This result suggests that fitness would be a well-suited criterion to determine the
topology inference power by the models. However, this is a very strong statement that
has to be confirmed or rejected. Therefore, to corroborate this result in the light of
possible parameter-overfitting, the individual-errors of every optimization run was
calculated by discretizing every resultant weight matrix of the optimization runs to a
ternary distribution according to the following mapping: [-5,-1] -> -1 , [-1,1] -> 0,
[1,5] -> 1. It is important to notice that even for this discretization, the goal network is
just one in 3N^2 = 19 683 possible solutions. In this way, information was obtained to
correlate individual fitness (MSE) with individual quantification of errors (individualerrors, from here on).
Since MSE and individual-errors distributions were not normal, a parameter free test
(Spearman) was performed to correlate these two variables along every model group.
In figure 4.9, the bars depict the Spearman correlation between the fitness, the
individual-errors and their respective p-values (shown by the lines with second y axis)
for each simulation interval and model respectively, after filtering for the 2 standard
deviations.
90
0.8
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
0.8
0.8
0.6
0.6
0.4
0.4
0.6
0.4
0.2
0.2
0.2
0.0
0
CTRNN-NP
0
-0.2
-0.4
0.8
0.2 0.3 0.5 0.6 0.75 1
correlation
1.5
2
3
Periods
p-values
0.6
0.4
0.2
0.0
0.2 0.3 0.5 0.6 0.75 1
correlation
1.5
2
3
Periods
p-values
TDRNN-P
0.6
0.4
0.2
Periods
0
-0.4
1
0.8
0.6
0.4
0.4
0.2
0.2
0.0
0
-0.2 -0.2
-0.4 -0.4
0.8
TDRNN
0.6
-0.2
-0.2 -0.2
-0.4 -0.4
0.8
CTRNN-P
0.2 0.3 0.5 0.6 0.75 1
correlation
1.5
p-values
2
3
Periods
0.2 0.3 0.5 0.6 0.75 1
correlation
1.5
p-values
2
3
-0.2
-0.4
1.0
0.8
0.6
0.4
0.2
0.0
-0.2
-0.4
Figure 4. 9 Spearman correlation between individual-run-errors and their respective fitness
(MSE) for the 4 models along the 9 optimized time intervals. Corresponding p-values are
show as thin red lines.
The correlation between individual errors and fitness is weak and varies with the data
structure and the model equation used to represent the GRN. For the simulation time
interval 0.5<T< 0.75, where T denotes the time in multiples of the oscillation period,
the MSE median was significantly lower (in the so called easy to fit region) for all
models with a weak, yet significant correlation between fitness and errors. Such a
correlation was strongest (Spearman ≈ 0.5, p < 0.05) in the region 0.2<T<0.3 for all
models. Optimizing for one oscillation period (T=1) it is found an anti-correlation
between fitness and errors in all models. Notice in figure 3.11 the strong change in
correlation between 0.75<T<1 for the CTRNN-NP model. In contrast to the TDRNN
model, the CTRNN model exhibits again a strong correlation for the remainder of the
MSE intervals (2 <T<3). This result rejects the suggestion that fitness could be a good
indicator of inference power.
To explain the existence of these 3 MSE zones, the dynamics of every optimization
run was analyzed for a longer period of time (T=7.2). In the upper part of figure 4.10,
this is exemplified by plotting the dynamics of three different period fractions from
the TDRNN-NP model representing these 3 regions. To measure, in some extent, the
stability of every solution, the instantaneous-fitness was calculated along this time
interval for every run. At the lower part of figure 4.10 is depicted the so called
instantaneous-fitness along the dynamics of every run of the respective dynamics
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
91
from the upper part. Here, the mean of the instantaneous-fitness of every model is
represented by a thicker black line and the standard deviation by a dashed red line.
The optimized interval is depicted by a vertical black dashed line.
20
10
Dynamics
5
10
5
0
0
0
-10
-5
-5
-20
0 0.2
2 0
Fitness
Periods
1
0.5
0
2
-10
0
1.5
1.5
1
1
0.5
0.5
0
0
0 0.2
0.75
Periods
Periods
0.75
Periods
2
0
0
2
Periods
4
2
Periods
4
Figure 4.10 Long term simulation dynamics of the TDRNN-NP model inferred from three
different time intervals (from 0 to the vertical dashed line).On the left side panels over-fitting is
present. On the middle panels de-synchronization occurs for the majority of the models after
the optimized interval. In the right side panels a stable long term fitting is observed for the
majority of the models. However, some failures to fit the amount of data exist.
While all solutions fit the data for 0.2<T<0.33 (figure 3.12 left panels), only few
solutions show the desired oscillatory behavior for long times. This could be
corroborated with the instantaneous fitness in the left lower panel of figure 4.10.
Simulating and fitting the system for 0.5<T<0.75 (figure 3.12 middle panels) all
solutions show stable oscillatory behavior, yet with different oscillation periods, as
they start to desynchronize for long times. Again, here it is easy to corroborate with
the respective (lower panel) graph of the instantaneous-fitness, where after the
optimized period a broad spectrum of similar frequencies is depicted. Finally, for T>1
(figure 4.10 right panels), two solution groups were found: one group of solutions
showed almost perfect agreement between model simulation and synthetic data, with
a stable long term instantaneous-fitness while the second group showed significant
discrepancies in the inferred amplitude and frequency of the oscillations also showing
a very poor instantaneous fitness.
92
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Taken together, all these results demonstrate that the TDRNN model could infer the
goal topology from as little information as just one third of the cyclic repressilator
model with just one error, approx. 90% of accuracy of correctly predicted weighted
and directed edges, outperforming the CTRNN model in terms of congruence with the
goal network.
4.1.3 Robustness against noise
To determine the influence of measurement errors on the reverse engineering process,
I added Gaussian distributed noise to the time series data. I equidistantly sampled the
repressilator sine functions at ten time points over a time interval of two cycles and
added noise with a standard deviation σ = sI , where I and s denote the amplitude of
the sine wave and a proportionality factor, respectively. The latter is used to define
the noise strength in subsequent experiments. To assess the performance of different
interpolation approaches under the influence of noise, I interpolated the sampled data
using a linear and cubic spline interpolation for s set as 20%, 30%, 40%, and 50%,
respectively. The new datasets were then reverse engineered to investigate the
relationships between measurement noise, data interpolation, the model functions and
the use of the parameter pruning resulting in 8 models under four noise conditions in
total. On figure 4.11 are the boxplots from the distributions of the fitness of these 8
groups.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
0
10
Fitness (log MSE)
0
10
TDRNN-P
TDRNN-NP
Fitness (log MSE)
Fitness (log MSE)
CTRNN-P-L
0
10
0
10
20 30 40 50
% of Noise
20 30 40 50
% of Noise
CTRNN-NP-S
CTRNN-P-S
TDRNN-NP-S
TDRNN-P-S
0
10
0
10
0
10
20 30 40 50
% of Noise
20 30 40 50
% of Noise
Fitness (log MSE)
20 30 40 50
% of Noise
Fitness (log MSE)
20 30 40 50
% of Noise
Fitness (log MSE)
Fitness (log MSE)
Fitness (log MSE)
CTRNN-NP-L
93
0
10
20 30 40 50
% of Noise
20 30 40 50
% of Noise
Figure 4.11 Data fits comparison among the two models (TDRNN and CTRNN), pruning (P)
and not pruning and using linear (L) or spline (S) interpolation giving a total of eight groups.
Distribution outliers are represented by a “+” sign.
The boxplots of the fitness from the optimization runs on figure 4.11 show that the
TDRNN models have less outliers (+ signs) than the CTRNN models, meaning that
the TDRNN models fail less often to fit the data. Comparing between the intergroup
medians, I found in all 8 groups, that the more the noise the worst (higher MSE) is the
fitting of the data. On the other hand, comparing corresponding intra group medians
among models, the TDRNN has always a better fitness than the CTRNN model.
Moreover, without the outliers, the distributions of the fitness (MSE) in all the
TDRNN model groups follow a normal distribution while the respective fitness
distributions from the CTRNN do not. Therefore, since the CTRNN shows too many
problems to fit the data, in order to compare among the media of both models and
interpolation approaches, a double filtering was applied.
Fitness
(MSE)
0.8
0.6
0.4
50%
40%
30%
TD_PR_L
20%
TD_NP_L
CT_PR_L
CT_NP_L
TD_PR_SP
TD_NP_SP
CT_PR_SP
0
CT_NP_SP
0.2
Noise
Models
Figure 4. 12 Inference power comparison among the eight
groups. Linear spline interpolation performs better that spline
interpolation for higher noise quantities
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
94
The first filter was to split those clear failure fittings with MSE >1, and the second
was the previously used +/- 2 standard deviations. The result of the means of this
double filtering is on figure 4.12.
On figure 4.12, it is clearly shown that besides the model utilized, in all cases the
models can better fit the data when using linear interpolation instead of the cubic
spline interpolation. This is particularly true for noise quantities ≥ 30%. This result is
of significance since on the reverse engineering area this is an open issue (Bar-Joseph,
2004; Bar-Joseph, et al., 2004). On the normal reverse engineering workflow
introduced in this thesis, the more important result is the calculus of the robust
parameters that determine the network topology. In this case, the comparison of this
robust parameters and its comparison to the goal network is plotted in figure 4.13 by
the means of errors as mismatches between inferred topologies and goal network.
CT_NP_L
CT_NP_S
6
CT_P_S
TD_NP_L
TD_NP_S
TD_P_L
E rrors
CT_P_L
4
2
0
20%
30%
40%
50%
Noise %
TD_P_S
Figure 4. 13 Inference power comparison among the two models (TDRNN and CTRNN),
pruning (P) and not pruning and using linear (L) or spline (S) interpolation giving a total of
eight groups, under different strength of Gaussian noise.
The direct result from the robust parameters calculation depicted in figure 4.13 shows
that the TDRNN is by far more robust against noise than the CTRNN model. Actually
the TDRNN models using linear (L) interpolation present no errors regardless of the
quantity of noise s in the data and the use or not of the pruning function (NP and P).
The same TDRNN models (NP and P) using spline interpolation showed one error
each for s ≥ 20 %. Interestingly the CTRNN model using spline interpolation and not
using the pruning function, showed no errors regardless of the presence of noise. By
contrast, the rest of CTRNN models present several errors for all quantities of noise.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
95
Finally, analogous to the analysis performed in the length of data section, individualerrors per run were calculated to correlate the fitness (MSE) of the runs with their
respective individual-errors. Again, this was achieved through the ternary
discretization of individual run’s matrix of weights and comparison to the goal
network.
After filtering for 2 standard deviation ≥ intra group mean of MSE, the CTRNN
model still has no normal distributions of its MSE runs. Therefore, Spearman’s (none
parametric) correlation was performed over the filtered (to avoid for spurious
correlations) MSEs of the models and the individual-errors of every run. The results
are plotted on figure 4.14.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
96
0.8
0.8
0.8
0.6
0.6
0.6
0.8
0.4
0.4
0.4
0.6
0.2
0.2
0.2
0.4
0
0.2
CTRNN-NP-L
0
0.0
20
p-values
0.8
30
40
Spearman
-0.2
50
Poly. (p-values)
CTRNN-NP-S
20
p-values
40
Spearman
0.8 0.8
CTRNN-P-L
30
50
1.0
0.0
Poly. (p-values)
0.8
CTRNN-P-S
0.6
0.6 0.6
0.6
0.4
0.4 0.4
0.4
0.2
0.2 0.2
0.2
0.0
0.0
0
20
p-values
0.8
30
40
Spearman
50
Poly. (p-values)
TDRNN-NP-L
1.0
0.8
0.8
0.6
0.8
0.6
0.4
0.6
0.4
0.2
0.4
0.3
-0.2
20
30
40
50
-0.7
p-values
0.8
0.6
0.4
0.2
0
-0.2
-0.4
0
Spearman
0.2
0
0.0
-0.2
Poly. (p-values)
0.8
TDRNN-P-L
20
p-values
0.6
p-values
30
40
Spearman
50
Poly. (p-values)
TDRNN-NP-S
1.0
0.2
20
30
40
Spearman
0.8
50
Poly. (p-values)
0.8
TDRNN-P-S
0.6
0.6
0.4
0.4
20
p-values
30
Spearman
40
50
0.2
0.0
Poly. (p-values)
0.4
0.2
0.2
0
-0.2
p-values
0.0
20
30
Spearman
40
50
0.0
Poly. (p-values)
Figure 4.14 Spearman correlation between the individual errors and individual fitness.
Comparison among the eight models groups under different strength of noise, correlation
trends are in dashed lines.
After filtering for outliers, figure 3.16 clearly shows that those models showing no or
a weak correlation between their MSE distribution and individual errors, are exactly
those with no errors encountered by the robust parameter calculation technique
(TDRNN NP-L, TDRNN P-L and CTRNN NP-S) depicted in figure 3.15. This result
could be seen counterintuitive at first glance, but bear in mind that since the errors are
of discrete nature and the fitness (MSE) of continuous nature, therefore small
variations on the fitness (MSE) do not correlate to the errors because they could have
no variation at all (zero error zones). Therefore, only strong correlations with low p-
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
97
values are significant and positive ones are associated to fitting problems while
negative ones to over fitting problems. However, usually one could not know the goal
topology in advance and these results are useful just to identify that fitness alone is a
bad indicator of the inference power of a model.
4.1.4 Robustness against incomplete information: Clustering
improves the standard reverse engineering task, quantitatively and
qualitatively
On this section, the model is evaluated for the common situation when the network
under consideration is actually larger than the number of genes selected for reverse
engineering (see Methods, data selection). In order to elucidate the effect of
incomplete information on the modelling procedure optimizations using three-node
CTRNN and TDRNN models were performed, while only the fitness in one or two of
the three nodes was measured.
The distributions of the fitness (MSE) and the individual-errors for this experimental
set-up are plotted in figure 4.15 trough boxplots. As expected, (figure 4.15 lower
panels) an increase in the number of individual-errors was found for both models with
a reduction of information. For the one-node-case all four models (TDRNN, CTRNN
both using and not pruning) were unable to properly infer the goal network, showing
more than four out of nine possible errors. Consistent with this observation, a
reduction in the number of nodes to be optimized increased the computed fitness as
shown in upper panels of figure 4.15 while they decreased in the sense of the MSE.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
98
MSE
0.1
0.05
0
CT-NP-1 2
3 CT-P-1 2
3 TD-NP-1 2
Models
3 TD-P-1 2
3
3 CT-P-1 2
3 TD-NP-1 2
Models
3 TD-P-1 2
3
8
Errors
6
4
2
0
CT-NP-1 2
Figure 4.15 Models fits comparison between the TDRNN and CTRNN using (P) and not (NP)
the pruning function, along different proportions of incomplete information represented by the
number of nodes optimized: 1 = two nodes not optimized, 2 = one node not optimized and 3
all nodes were optimized.
Concerning the calculus of the number of errors through the identification of the
robust parameters approach, a similar tendency was found, here depicted in figure
4.16 Even though the TDRNN models have significant less number of errors for the
case of two optimized nodes, in general all models failed when just one out of three
nodes were optimized.
As mentioned, this case of incomplete information occurs very often in the RE of
GRN area. Therefore, it would be of high interest for the community to increase the
inference power of any model for such a situation. Hence, here I performed an
additional analysis with the objective of improving the inference power of my model.
In the two previous sections, it was shown that different network topologies are
inferred along the reverse engineering process (Figures 4.10). To systematically
analyze these distributions, the total sets of different parameter solutions were
clustered. For this purpose, the matrix of weights of every optimization run was
represented as a vector.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
CTRNN_P
CTRNN_NP
TDRNN_P
TDRNN_NP
6
5
Errors
99
4
3
2
1
0
1
2
3
Nodes
Figure 4. 16 Inference power along different quantities of missing
information. 3 = no missing information
The next step was to identify the ideal number of clusters to split the entire set of
solutions. Therefore, hierarchical and SOM clustering (see chapter 3 Methods) were
applied over the sets of 50 vectors per model and where determined five as the
number of naturally formed clusters. Then I used the k-means algorithm to split the 50
optimization runs on each of the four models. Finally, I recalculated the robust
parameters to obtain the adjacency matrix for each cluster.
The first and more important result of this section appears for the analysis of the
clusters of the different models when the optimization of just one node was
performed. The following results are related only to that case. Additionally, since five
different possible solutions per model were obtained, I needed a way to differentiate
among them. Therefore, the inference power of every cluster was calculated as the
ratio between its size and its MSE mean fitness. The upper panels (a-e) of figure 4.17
depict the five clusters of the TDRNN-NP model, the mean (over cluster row in
colour scale) and the standard deviation (over the mean row in grey scale) of every
cluster and their inference index. The respective weighted directed graph is obtained
from the robust parameters calculations from the last two parameters (see chapter 3,
Methods) and it is placed below their respective cluster. On figure 4.17, on the lower
panels (f-j) are depicted the analogous results for the CTRNN-NP model. The colour
scale of every cluster and their mean is depicted in figure 4.17 k.
100
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
L)
k
)
W 1,2
W 2,2
W 1,1
W 2,1
W1,3
W 3,2
W3,1
W2,3
W 3,3
Figure 4.17 On top of a, b ,c ,d and e are clusters representing groups of weight parameter
solutions between the three nodes from the TDRNN-NP model. Every row is a vector
representation of the matrix of weights (W i,j) from each of the 50-optimization runs from this
model. At the top of ever cluster are the columns heads for mapping positions (W ij) of these
vectors into the matrices of weights, below these column heads are the standard deviation
(grey scale) and the mean value (color scale) of every column. On k) is placed the weights
color scale and below in l) is a scheme with the map of weights positions representing the
repressilator (blue dashed lines does should not exist, showed just for mapping purposes).
With the mean and standard deviation, the z-score is calculated and those robust parameters
(z-score > 2) determine the edges on the lower graphs. Networks b) and e) are closely
related. Those networks and network a) are topologically equivalent solutions to the
repressilator, with the rotation direction being the only difference between them. To
differentiate between the four architectures encountered (a, b + e, c and d) we calculated the
ratio between number of elements and mean fitness per cluster and named it the cluster
index. Those with the highest index are the two repressilator-like architectures. On f, g, h, i
and j, are the analogous clusters from the CTRNN model and below them, their respective
inferred networks. Here i and j are related solutions. The network on a, is the counter wise
repressilator solution, despite the cyclic behavior is incomplete
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
101
While performing a visual inspection of the formed clusters, an unexpected feature
emerged. For the entire repressilator study, a counter clockwise architecture as the
goal network was considered (see figure 4.1a). However, while optimizing just one of
three nodes this restriction banishes because only the sinusoidal dynamics of one node
is important. In this way, the equivalent cyclic repressilator architecture functioning
clockwise is also a valid topology which is depicted in figure 4.17-L. Taking this
second architecture into account, it is easier to explain the fact that the three clusters
with the largest index (> 1000 units) from the TDRNN-NP model (clusters a b and e
from figure 4.17 upper panel), have an equivalent repressilator graph among them.
Cluster-graph a is the counter clockwise repressilator architecture while clustersgraph b and e represent the second valid clockwise repressilator architecture depicted
in 4.17-L.
Performing an analogous analysis for the CTRNN-NP model leads to the graphs f-j
from figure 4.17. However, notice that here are only two clusters with an index >
1000 units and that the model failed to find the cyclic topology of the counter clock
wise repressilator (figure 4.17f). For the other two models, as well as for the previous
ones, the results are summarized on table 4.2. However, in table 4.2, I additionally
compared some other properties of the directed graphs derived from the clusters as the
number of false positives, false negatives and if a cyclic topology of any repressilator
were found. As it can be seen, in all four models the cluster-graphs with the higher
index (bold fonts) are those related to any of the two repressilator architectures.
Taking only these solutions into account, all the models decrease their number of
errors dramatically.
102
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Table 4. 2 Comparison among the different clusters and networks from the two different
models TDRNN and CTRNN using (P) and not (NP) the pruning function
MODELS
CTRNN_NP
CTRNN_P
TDRNN_NP
TDRNN_P
Errors
F. positives
F. negatives
Cluster size
MSE mean
Cluster index
CYCLE
Errors
F. positives
F. negatives
Cluster size
MSE mean
Cluster index
CYCLE
Errors
F. positives
F. negatives
Cluster size
MSE mean
Cluster index
CYCLE
Errors
F. positives
F. negatives
Cluster size
MSE mean
Cluster index
CYCLE
Clust. 1
2
0
2
20
0.00
4163.59
NO
3
0
3
14
0.01
1828.78
YES
1
0
1
25
0.01
4726.47
YES
4
1
3
5
0.01
639.78
YES
Clust. 2
2
1
1
2
0.00
774.29
NO
4
1
3
10
0.05
199.85
NO
2
1
1
8
0.00
3732.65
YES
2
0
2
15
0.01
2451.19
YES
Clust. 3
5
1
4
9
0.01
1010.11
NO
4
1
3
7
0.05
129.00
NO
7
3
4
4
0.01
268.00
YES
5
1
4
12
0.05
246.57
NO
Clust. 4
2
2
0
2
0.00
687.99
YES
8
3
5
3
0.04
78.17
YES
5
1
4
2
0.01
134.39
NO
5
1
4
9
0.05
181.61
NO
Clust. 5
1
0
1
17
0.00
3710.31
YES
3
0
3
16
0.01
2373.65
YES
2
0
2
11
0.01
1472.27
YES
2
0
2
9
0.01
1137.29
6
YES
F positives = false positives
F negatives = false negatives
CYCLE = cyclic inferred network
Summarizing, applying the cluster analysis, a decrease in the error incidences from 4
to 1.5 was found for both CTRNN-NP and TDRNN-NP models. However, only for
the TDRNN-NP model was found three different network topology solutions with
oscillatory behavior. These clusters with the highest inference power constitute the
topologically equivalent repressilator architectures with clockwise or counterclockwise cyclic repression (see figure 4.17 b, e and a respectively).
Clustering solutions of the CTRNN-P and TDRNN-P models decreased the error
incidences from 6 and 5 (see figure 4.16) to 3, respectively (see table 4.2). Both
models were able to find the two repressilator architectures. Repeating the same
cluster analysis for reverse engineering with two nodes all models succeeded to infer
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
103
the desired repressilator topology without errors. This demonstrates clearly that the
parameter clustering approach improves the model’s inference power.
4.2 The yeast cell cycle
To assess the predictive power of the TDRNN as a gene regulatory network model
and the inference power of the clustering extended reverse engineering workflow for
biological systems, in this section, I compared the performance of the TDRNN with
that of a CTRNN and a Dynamic Bayesian Network to infer the well studied (Chu, et
al., 1998; Futcher, 2002; Gavin, et al., 2006; Guelzim, et al., 2002; Harrison, et al.,
2007; Ihmels, et al., 2002; Krogan, et al., 2006; Murray and Beckmann, 2007; Tsai
and Lu, 2005) transcription-signal transduction cell cycle network of Saccharomyces
cerevisiae based on experimental data.
The goal network
On their work, Li et al. (Li, et al., 2004) proposed a cell cycle network model of
Saccharomyces cerevisiae having 11 nodes that comprise 18 genes, proteins and
black boxes. The functional network elements are divided into 4 groups: cyclines (Cln
1,2 and 3), inhibitors degraders and competitors of the cycline cdc28 (Sic1, Cdh1,
Cdc20/14), transcription factors (SBF, MBF, Mcm1/SFF, Swi5), check points and self
repressions (cell size, DNA replication), see figure 4.18a. There are 34 interactions,
out of which 19 are positive and 15 negative. Note that five negative interactions are
of unknown nature and were added in order to make the network functional in logical
terms.
The data source, selection and quality control
In this section, the gene expression kinetics of the yeast cell cycle from the alphafactor data set was used (Spellman et al., 1998). Utilizing four different techniques to
synchronize the yeast colonies they generate four data sets named; cdc28, cdc15,
alpha factor, and elutriation. Unfortunately, other works (Fellenberg, et al., 2001) has
shown that desynchronization occurs on the cdc15 and elutriation data sets; therefore
104
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
we do not consider them. After analyzing the left two data sets, we discarded the
cdc28 data set because the high frequency of missed data points on our selected
genes. Therefore, I end up with just the alfa factor data set. After performing the
quality control described on methods, no particular issue was detected for the selected
data set.
However, inferring the cell cycle network proposed by Li et al from this
transcriptomic data is challenging, as only four of the eleven network nodes are
transcription factors and it includes 2 dimers (SBF, MBF) and 5 nodes that are
represented by 2 proteins: Clb 1/2, Clb5/6, Mcm1/SFF, Cdc14/20 and Cln1/2.
Fortunately, the genes encoding the dimers and the redundant proteins exhibit similar
expression kinetics so that I represented these nodes by the mean expression of their
respective genes.
Data interpolation
I tested two different interpolations approaches for this data set: linear interpolation
and a B-spline interpolation method suggested by Bar-Joseph (see Methods) for the
continuous representation of gene expression.
Models
The data was fitted using the TDRNN and the CTRNN models choosing the
parameter ranges for
[τ ,θ,ϑ ]
to [10 − 55,0 − 1,0 − 3.5]. For the TDRNN, an
optimization time delay range equivalent to 10 minutes was chosen as a biological
compromise between the fast signal-transduction and the slow transcriptional
responses.
4.2.1 TDRNN shows superior inference and predictive power than
previous models on experimental data
Here, I compared the TDRNN performance on the cell cycle data with CTRNN and
Dynamic Bayesian Networks (DBN) as described in chapter 3. The quantitative
comparison of the different solutions was performed by computing the adjacency
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
105
matrix for each group of model solutions and calculating the cost of transforming the
resulting directed graph of every model group into the goal network.
As an objective transformation cost function, a directed-weighted version of the graph
edit distance algorithm (GED) was used. This algorithm calculates the cost to convert
one graph into another by changing edge weights and/or directions. Additionally, it
accounts for shortcuts (the deletions-insertions of nodes) from the semantic of
directed, weighted graphs. The results for the GED together with other graph
measures are depicted in table 4.3
Table 4. 3 Mean square errors (MSE) and graph edit distance from inferred networks using
the CTRNN, Dynamic Bayesian Network and TDRNN models. The MSE is averaged over 50
optimization runs. The columns two and three denote the number of nodes and edges for
which robust interactions were inferred. Columns four and five show the number of correctly
and falsely predicted interactions, which column 6 shows the GED costs. The rows with label
Cluster 1-4 show the MSE and the GED for the individually clustered solutions from 50
independent optimization runs for the TDRNN-NP model. The Bootstrapping results for model
are given in the bottom row
Model
MSE Nodes Edges Positives FP GED Index
CTRNN
0.98
3
2
0
2 134
Dynamic Bayesian network NA
11
34
10
24 119
TDRNN B-Spline
0.71
9
9
2
7 107
TDRNN-P
0.36 10
18
8
10 80
TDRNN-NP
0.37 11
19
8
11 68
Cluster 1
0.36 11
21
9
12 62
50
Cluster 2
0.36 11
26
15
11 41
28
Cluster 3
0.37 11
24
9
15 61
22
Cluster 4
0.39 11
25
11
14 70
20
Bootstrap
GED Mean =99, GED Standard deviation = 147
Nodes = number of nodes with at least one edge Positives = number of correct inferred
correlations among two nodes F P = false positive edges
GED = graph edit distance
cost
According to this analysis, the TDRNN-NP model using linear interpolation has the
best inference power of all models having the lowest GED costs. The results depend
crucially on data interpolation where this model shows also the best predictive power
exemplified in figure 4.18 and to a lesser extent on parameter pruning, changing the
GED cost by 64% and 23%, respectively.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
4
3
2
1
0
-1
-2
-3
-4
3
Expression
Expression
CLN3
1
0
8
6
4
2
0
-2
-4
-6
41
61
81
Time (min)
-1
-3
-1.5
101
1
21
5
4
3
2
1
0
-1
-2
-3
41
61
81
Time (min)
101
Expression
21
3
41
61
81
Time (min)
101
6
0
-1
-2
-3
1
21
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
41
61
81
Time (min)
21
41
61
81
Time (min)
21
3
Expression
Expression
2
0
101
1
21
6
4
2
0
-2
-4
-6
-8
SWI-5
1
CDH-1
CLB-5,6
4
-6
1
101
101
-4
Expression
1
41
61
81
Time (min)
-2
3
2
1
0
-1
-2
-3
-4
-5
CDC-14,20
2
21
41
61
81
Time (min)
101
CLB-1,2
Expression
1
1
SIC-1
Expression
CLN-1,2
0
-2
Expression
21
MBF
0.5
-0.5
-1
1
Expression
1
SBF
2
Expression
106
41
61
81
Time (min)
101
1
21
41
61
81
Time (min)
101
MCM1
2
1
0
-1
-2
1
21
41
61
81
Time (min)
101
1
21
41
61
81
Time (min)
101
Figure 4.18 Data fitting example from one of the TDRNN runs. In straight lines are the
original linearly interpolated data; in smooth lines are the approximation of every node of
one model for their respective gene expression. The global fitness of this average run was
MSE= 0.36
Notice that on figure 4.18, the approximation of most of the gene dynamics is
acceptable. Only for the MCM1 gene, the approximation presents some discrepancy
to the data (straight line). The reason for this behavior is that the data for this gene
presents some missing data point. Therefore its pattern appears almost randomly
changing from active to inactive. This is a limitation from the data set and therefore, it
was not considered necessary to improve its fitting.
4.2.2 Bootstrapping validation
To validate the TDRNN-NP results for the yeast data, a bootstrapping test was
performed by randomly shuffling the order of the microarray time series data 50 times
and repeating the entire workflow. Then, the GED cost of each of the 50 adjacency
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
107
matrices obtained was calculated. The resulting GED costs were normally distributed
as expected (see figure 4.19). Additionally, the Lillie test of normality was performed.
The mean and standard deviations of the bootstrapped optimization runs are shown on
the bottom row in table 4.3. Notice that the mean of the GED cost from this
bootstrapping deviates more than 2 standard deviations in comparison with the
TDRNN-NP model. This confirms that the results from our TDRNN-NP model were
not obtained by chance which further validates the correlation between the inferred
network topology and the proposed cell cycle network.
12
Frequency
10
8
6
4
2
0
70
78
86
94
102
GED
110
118 More
Figure 4.19 Bootstrap data histogram. The distribution of the
cost to transform networks, inferred from randomizing the
original data, into the goal network is normal with mean = 99
edition units.
4.2.3 Clustering improves the RE process with real data
After performing the clustering approach described in the section 3.1.4, I found that
the two clusters with the highest index (clusters 1 and 2) decreased the GED cost from
68 to 62 and 41, respectively (see table 4.3 and fig. 4.20b). The GED cost from cluster
2 deviates by more than four standard deviations from the mean of the bootstrapping
test, i.e. the result is unlikely to be obtained by chance.
108
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Figure 4.20 On the left side, the original Yeast cell cycle network from (Li et al., 2004), where
blue and red arrows mean activation and inhibition, respectively. The indispensable edges
according to the work from Stoll et al. are on thicker lines. On the right side, is the network
topology inferred by our TDRNN model, (cluster2, of the yeast results on table 2). Red and
blue thicker lines (8) are correctly inferring directed weighted edges, green lines (7) denote
matches with reported correlations between two nodes on the goal network (direction and/or
weight are not properly inferred on this edges) and grey dashed lines are false positives (11).
Seven of the 8 properly inferred weighted directed edges, belongs to the 13 indispensable
ones reported by Stoll while none of the misdirected/weighted falls into this category.
Comparing the inferred network from cluster two (figure 4.17b) with the cell cycle
goal network (figure 4.20a), I found that 7 out of 8 weighted and directed edges were
predicted correctly according to the 13 edges defined by Stoll (Stoll and Rougemont,
2006) as being the only necessary interactions for this yeast cell cycle circuit. The low
graph editing costs of the best inferred network, i.e. cluster 2, demonstrates that the
TDRNN model together with the clustering of reverse engineered solutions is able to
extract biologically meaningful information from a combined protein signalling and
gene regulation network by just using one experimental time series (44% [15/34] of
the total correlations, or 54% (7/13), if considering the indispensable weighted
directed edges).
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
109
4.3 Reverse engineering of keratinocyte-fibroblast
communication
In this section, I present a case study of an unknown network. This scenario is a real
situation where there is little information about the structure of the function one is
interested in to understand its dynamics through reverse engineering. Along this case,
the assumptions, limitations and improvement opportunities that offer such
experimental scenario will be explained. Probably, the more important gain from this
case scenario lies in making the experience of how to direct a data-driven
experimental setup as well as how to avoid certain problems in the iterative
experiment-modeling process.
The unknown goal network
The idea behind this experimental setup was to study the communication between two
cell lines, keratinocytes and fibroblast. This interaction is important for processes like
skin wound healing and some authors (Birchmeier, et al., 2003) have even suggested
that it is related to cancer by the means of certain analogies. In particular, the process
here studied is related to cell migration and its cancer analogous: metastasis.
To study this interaction, DNA microarray experiments at several time points upon
hepatocyte growth factor (HGF) stimulation was performed by a cooperation partner
group (Axel Szabowski), to obtain the gene expression kinetics from heterogonous
co-cultures containing primary human keratinocytes and murine cjun-deficient
fibroblasts. The latter were chosen trying to discriminate between human and murine
mRNA, based on species-specific sequences and to provide an HGF-free background.
This is a strong assumption; therefore a detailed quality control analysis will be
explained for this data set.
In the global experimental setup keratinocytes were stimulated with HGF, which
induces both proliferation and migration (Birchmeier et al, 2003). Three additional
experiments using keratinocyte growth factor (FGF-7), granulocyte–macrophage
110
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
colony-stimulating factor (GM-CSF) and stromal derived factor-1 (SDF-1) as stimuli
were conducted, all of them inducing cell proliferation, but not migration (Florin, et
al., 2004). The particular experimental details are described below in order to explain
the analysis and the obtained results, but the experiments where performed by the
collaboration partner group.
Cell culture
Normal human skin keratinocytes and dermal fibroblasts (HDF) were derived from
adult skin (Stark, et al., 1999). HDF obtained from the outgrowth of explant cultures
were grown in Dulbecco’s modified Eagle’s medium (DMEM; Bio Whittaker)
supplemented with 10% fetal calf serum (FCS), and cells from passages 4 to 8 were
used. Mouse wild-type and cjun-/- fibroblasts were isolated from mouse embryos and
immortalized according to the 3T3 protocol (Schreiber, et al., 1999) and used together
with HDF as feeder cells. Normal human skin keratinocytes were plated on Xirradiated feeder cells (HDFi, 70 Gy8; MEFi, 20 Gy) in FAD medium (DMEM/Hams
F12 3:1) with 100 U/ml penicillin, 50 mg/ml streptomycin and supplemented with 5%
FCS, 5 mg/ml insulin, 0.1 ng/ml recombinant human EGF, 0.1 nM cholera toxin, 0.1
nM adenine and 0.4 mg/ml hydrocortisone (Sigma). For expression profiling, total
RNA of co-cultured cells was isolated 1, 2, 3, 4, 6 and 8 h after stimulation with
recombinant human cytokines (10 ng/ml HGF, 10 ng/ml GM-CSF, 10 ng/ml FGF-7
or 10 ng/ml SDF-1; all obtained from R&D Systems).
Migration assay
Immortalized human keratinocytes (HaCaT cells) were cultured in monoculture with
DMEM (10% FCS, 100 U/ml penicillin, 100 mg/ml streptomycin). Subsequently, the
cell monolayer was damaged with a ‘scratch’ using a pipette tip and the cells were
8
In Dosimetry, which is a scientific subspecialty in the fields of health physics and
medical physics that is focused on the calculation of internal and external doses from
ionizing radiation, the absorbed dose is reported in gray (Gy) for the matter or sieverts
(Sv) for biological tissue, where 1 Gy or 1 Sv is equal to 1 joule per kilogram.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
111
treated with 5 mg/ml mitomycin c (Sigma-Aldrich) 3 h before the stimulation. The
cells were stimulated at the indicated time points and periods with cytokines and/or
inhibitors: 10 ng/ml HGF (R&D Systems), 1 ng/ml EGF (R&D Systems), 150 nM
tyrphostin AG1473 (Biomol) (EGFR inhibitor) or 1 mM GW2974 (Sigma) (EGF-R
inhibitor), 50 mM meloxicam (Biomol) (PTGS-2 inhibitor), 0.5 mM H-89
(Calbiochem) (PKA inhibitor) or 10 mM myristoylated PKI (14–22) amide, cellpermeable PKA inhibitor (Biomol), 200 mM 8-(4-chlorophenylthio) adenosine 30, 50
-cyclic mono-phosphate sodium salt (Sigma) (PKA activator) and incubated for
further 30 h. The Relative migratory activity was determined by measuring the
migration distance during the culture by using standard protocols.
In general, the experimental setup has three strong issues: the first is the fact that the
cells were not synchronized in order to perform the measurements of they
transcriptomic response after stimulation. This is crucial since the measurements are
done over populations of cells, and it should represent the average behavior of the
population. The other two are described on the next paragraphs.
Microarray data acquisition and analysis
Microarray measurements were recorded for four different stimuli from co-cultures,
namely HGF, FGF-7, SDF and GM-CSF. For each stimulus, within one experiment
six probes were taken (time points of 1, 2, 3, 4, 6 and 8 h after initial system
stimulation) and further analyzed. Total RNA was isolated, labeled and hybridized to
HG-U133-2plus (Affymetrix) according to the manufacturer’s protocol. Raw
microarray data were processed using the R environment (R Development Core
Team, 2007) and the Bioconductor toolbox (Gentleman, et al., 2004). The Probe
annotation
was
handled
with
the
Bioconductor
package
hgu133-plus2cdf
(Bioconductor Project). The Normalization was performed using variance
stabilization available in the Bioconductor package vsn (Huber, et al., 2002). The
gene fold expression was calculated according to the mean gene expression of two
control measurements of an uninduced system at 0 and 8 h.
112
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
The second important issue from that experimental set-up appears in this section,
where microarray measurements were performed over the mixture of co-cultured
cells: Immortalized Human keratinocytes, irradiated Human Fibroblasts and irradiated
Mouse Epidermal Fibroblast. Hence, cross hybridization occurs among different
mRNA expressed by these cells. Moreover, there is no information on how strong and
for which genes this cross hybridization could be.
Finally, the third important drawback of the experimental set-up resides on the lack of
measurements of the errors and/or noise associated to the previous two drawbacks. In
other words, there are no replications for the experiments because they are not
reproducible (Szabowski, personal communication). As mentioned, despite repetitions
could be expensive tedious and with low information content from an experimentalist
point of view, they are of central relevance to develop models of biological systems.
Despite these three drawbacks invalidate the data for any scientifically based result,
here will be shown as an example, what could be done with such data and some
suggestions about how to improve future works that could face similar problems.
Quality control
As described in Methods, the first step to start the reverse engineering is to asses the
quality of the data and to identify possible issues (Wilson and Miller, 2005). Figure
4.21 depicts the results of the preformed quality control of the keratinocyte arrays as
described in Methods. The interpretation in this case, is based on the following
explanations but, since these microarrays and quality control standards are thought for
a single homogeneous cell line experiments, biochemical interpretation will be added
when considered pertinent.
The figure is plotted from the bottom up, with the first chip at the base of the diagram
and the last chip at the top. This corresponds to the order of the samples depicted in
table 4.4.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
113
Table 4. 4 Microarray time series sampling and input strength
[µg/µl]
HGF
Chip number
Hours
b26
Control 8h
a1.
Control 1h
0.5
2.
HGF 1h
0.5
3.
HGF 2h
0.5
4.
HGF 3h
0.5
5.
HGF 4h
0.5
6.
HGF 6h
0.5
7.
HGF 8h
0.5
Figure 4.21 Quality control of the microarray data. Dotted
horizontal lines separate the plot into rows, one for each chip.
Dotted vertical lines provide a scale from -3 to 3 fold
expressions. Each row shows the percentage of responding
gene-probes in relation to the total gene probes in the chip, the
average background, the scale factors and GAPDH / β-actin
ratios for an individual chip.
The first indicator of the quality control pipeline is the average background; it is at the
left of figure 4.21. The variation among the different chips is bigger than 10%,
114
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
therefore they are colored red. This parameter is usually more related to experimental
problems, such as different concentration labeling or different efficiency of
hybridization cocktails. Here, a stronger variation is found at the chip number 5,
having a strong average background (67 units, compared to the 43 average of the two
referential 1 and 26 chips). Usually this parameter has to be analyzed together with
the rest of the quality control before obtaining false conclusions; however one has to
bear in mind the particularities associated to the mentioned issues. Therefore, this
strong background of the chip 5 will be analyzed later.
In figure 4.21 GAPDH 3’:5’ values are plotted as circles. According to Affymetrix,
they should be about 1. The obtained values suggest a misbalance of the different
transcripts sections (see methods section). However, no issues could be concluded
here.
β-actin, 3’:5’ ratios are plotted as triangles. Because this is a longer gene, the
recommendation is for the 3’:5’ ratios to be below 3; values below 3 are colored blue,
those above, would be red. This result indicates no early degradation issues of the
probes.
The percentage of present gene-probes is listed at the left side of the figure. The
variation among the different chips is less than 10%; therefore, they are colored blue.
However, it is important to notice that in general, the values are very low. This could
have two explanations. The first is that this is a normal result, because it is not
expected that all the genes are responding to the stimuli, instead just a small fraction
(around 33% of the total gene-probes). The second could be that the low percentage
of responsive probes is very likely an indicator of the mixture of cells. Though this
parameter is designed for homogeneous cell population measurements, the normal
interpretation does not apply. Instead, since probesets are flagged marginal or absent
when the PM values for that probeset are not considered to be significantly above the
MM probes, the number of mismatches could vary due to cross hybridization of the
mixture with genes from different cells.
Finally, the blue stripe in the graph represents the range where scale factors are within
3-fold of the mean for all chips. Scale factors are plotted as a line from the centerline
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
115
of the image. A line to the left corresponds to a downscaling, to the right, to an up
scaling. If any scale factors fall outside this ‘3-fold region’, they would be colored
red, otherwise they are blue. As shown in figure 4.21, there are no issues related to
large scaling factors.
Dimensionality reduction: Data selection and interpolation
As explained previously, data selection and interpolation is applied to reduce the
dimensionality of the data. The first step, data selection, is crucial to find the real
response of the biological system and no bias should be applied here towards a
particular set of genes. However, ideally, no genes balancing the overall system
should be missed. It is a difficult compromise. For this data set, it was better to rely
on the expertise of the experimentalist designer because the lack of controls could be
in some extent compensated by some assumptions from the area of interest. In this
sense, migration is the desired phenotypic response to be associated to a selected set
of genes. Therefore, genes were selected by an experimentalist expert, to amplify the
knowledge of their association to the migration process. Additionally, genes were
selected due to their expression profile. In this sense, the idea behind was to select
genes of three different kinds: early genes, possible target genes of the previous ones
and late response genes that could be target genes of the two first groups.
In this way, the selected genes were PTGS2, CEACAM1, ETS1, EGR1, JUNB, FOS,
PLAU, ITGAV, ITGB6, SERPINE, LAMA3, LAMC2 and additionally, I added
PLAUR in order to check for its possible role as an autrocrine feedback loop. The
expression profile of the selected genes is depicted in figure 4.22.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
116
1
2
3
4
6
8
3
2.5
2
1.5
1
0.5
0
-0.5 0
1
2
Time(h)
2
3
4
6
8
2
3
Expression
4
6
8
3
2.5
2
1.5
1
0.5
0
-0.5 0
1
3
4
6
8
4
6
8
Time(h)
3
2.5
2
1.5
1
0.5
0
-0.5 0
1
2
4
6
8
1
3
2
3
4
6
8
4
6
8
Time(h)
ITGAV
4
6
8
3
2.5
2
1.5
1
0.5
0
-0.5 0
1
2
3
Time(h)
CEACAM1
Expression
Expression
3
2
3
LAMA3
3
2.5
2
1.5
1
0.5
0
-0.5 0
PLAU
Expression
2
2
Time (h)
SERPINE1
1
1
Time(h)
Expression
Expression
2
Time(h)
3
2.5
2
1.5
1
0.5
0
-0.5 0
8
Time(h)
PLAUR
Expression
1
6
Expression
1
Time(h)
PTGS2
3
2.5
2
1.5
1
0.5
0
-0.5 0
4
ITGB6
3
2.5
2
1.5
1
0.5
0
-0.5 0
Expression
1
3
3
2.5
2
1.5
1
0.5
0
-0.5 0
Time(h)
ETS1
3
2.5
2
1.5
1
0.5
0
-0.5 0
FOS
Expression
JUNB
Expression
Expression
EGR1
3
2.5
2
1.5
1
0.5
0
-0.5 0
3
Time(h)
4
6
8
3
2.5
2
1.5
1
0.5
0
-0.5 0
1
2
3
4
6
8
Time(h)
Expression
LAMC2
3
2.5
2
1.5
1
0.5
0
-0.5 0
1
2
3
4
6
8
Time(h)
Figure 4.22 Selected data expression profile. Error bars does not represent experimental
repetitions, instead represent standard deviation among probe sets of every gene at every
sampling time point. Standard deviation represents the confidence of every selected gene
data.
The genes in figure 4.22 are ordered by their expression profile. From top to bottom,
there are from the early responding to the late responding genes. Here, it is important
to notice that the standard deviation bars does not represent repetitions, instead it
represents the variation of signal intensity for those genes which have more than one
probe-set. Hence, this standard deviation could be used just to measure the specificity
of every probe-set, but not to demonstrate the reproducibility of the experimental
setup.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
117
In this sense, genes without standard deviation bars as JUNB and FOS are
represented by just one probe on the microarray chip. Even though, these two genes
belong to an important family of genes with several homologous genes as the AP-1
system, there is just one probe for each member of the family. Therefore, one has to
take special care, because the possibilities for a cross hybridization with the respective
genes from the other cellular lines increase.
Genes like PTGS2 and PLAU exhibit a very small standard deviation among their
different probes; this shows that some genes do not have cross hybridization among
different cellular lines. Instead, their probe-sets are specific for them. By contrast, the
rest of the genes exhibit a large standard deviation among their probe-sets. This
situation is very likely due to cross hybridization.
The putative functionality of the selected genes is depicted on Figure 4.23. Notice that
the early responding genes: EGR1, JUNB, FOS and ETS1 from figure 4.22 are all TF.
Additionally, the apparently early responding genes from the upper panels of figure
4.22 (ITGV6 and LAMA3) are genes which encode for a cell receptor and do an extra
cellular kind of Laminin proteins respectively.
Stimuli
Integrins
•Complex formation
•migration
Cell adhesion
Cell signalling
•Laminins
•Adhesion
•migration
Plasminogen
inhibition
Early genes
•Inflammation
•Inhibition of anoikis
Figure 4.23 Thirteen selected genes, available information scheme.
118
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
The genes which maximum response – which happens after 2h (the middle panels of
Figure 4.22 )are the PTGS2, SERPINE1, ITGAV and PLAUR genes. The first two
genes, PTGS2 and SERPINE1, are encoding for inhibitory proteins associated to cell
signaling events, while the last two genes encode for cell receptors associated to the
migration process as depicted in Figure 4.23. Lastly, there are genes which expression
profile, in the lower panels of Figure 4.22 shows an increasing late response like
CEACAM1, LAMC2 and PLAU. Their cellular functionality varies: CEACAM1 is a
gene encoding for a protein receptor associated to processes as cell adhesion and
signaling; LAMC2 is another gene encoding for a Laminin protein associated to cellto cell adhesion and the migration process and PLAU is a gene encoding for the
excreted protein PLAU that could display an autocrine feedback loop for this process.
Concerning the interpolation process, since there is no information about the
associated noise, linear interpolation was utilized. However, two important aspects
need to be taken into account to configure the final data set to be engineered. The first
one is related to the quality control of Chip 5 and the strong background signal
expression at 4 hours. A careful visual inspection of the data depicted in figure 4.22
coincides that this data point is showing very likely an artificial inflexion in almost all
the selected genes. Inflexions in data are the means by which any correlation or multi
regression model could associate putative interactions among genes, therefore a
spurious inflexion could play a strong artificial role in the final result. By contrast,
notice that the gene FOS at the three hours point shows a second wave that could not
be associated to an experimental issue as demonstrated in the quality control and
therefore it may remain there. Hence, the decision was taken to discard the data for
the 4 hours point for the reverse engineering process as depicted in figure 4.24.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
3
Expression
2.5
2
1.5
1
0.5
0
-0.5
0
1
2
3
Time (h)
6
8
119
ETS1
JUNB
FOS
EGR1
CEACAM1
PLAU
PTGS2
LAMA3
LAMC2
ITGAV
ITGB6
SERPINE1
PLAUR
Figure 4.24 Linear interpolation of the expression profile of the
selected gene data. Due to quality control issues, data of 4h
were discarded. Interpolation data of the protein encoding
PTGS2 gene appears to be first in reacting to the stimuli
(stronger slope). The FOS transcription factor encoding gene
presents a plateau between 2 and 3 h.
The second issue that could rise from interpolating the selected data is that an
unexpected strong, early signal comes from some genes that do not belong to the early
genes as PTGS2. Notice in figure 4.24 that the strong slope from this gene could
direct the model to associate it as an initiator of the response. Hence, the importance
of an adequate sampling rises because an intermediary 30 minutes point may easily
show an even more abrupt response but at later time for this gene than the early genes.
Data fitting
As it can be seen in figure 4.24, the profile of the data to be fitted is relatively simple
compared to the previous two experiments from sections 4.1 and 4.2. Therefore, the
initial prescreening of the parameters showed that a much smaller population of only
100 individuals could be utilized to fit the data. Additionally, since the human cellular
life span and the response of the selected set of genes is larger than the respective
Yeast or E. coli of the two previous theoretical experiments, here a larger time delay
parameter has to be utilized. However, as it can be seen from the results of section
4.1.1, an even short increment in the time delay range could have undesired results
(see figure 4.2). Therefore, a maximum of 45 units’ equivalents to minutes from the
data set of figure 4.24 is utilized to fit this data. The selected range is of extreme
importance, especially if one considers that this experimental setup includes an
external input, because the fitting could be easily just driven by the external input and
120
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
when this is withdrawn, the selected functional module will withdraw its behavior,
neglecting that even without strong external signals genes interact with each other at a
basal level. Actually, this is the reason for the bias term and is the only source of
activity for the repressilator modeled in section 4.1.
As mentioned before, this experimental setup includes the addition of HGF as a firing
signal response. Therefore, here was used the equation (3.18):
τ
dYi N
= ∑ Wijσ (Y j ( t − δ i ) ) − θ j − ϑiYi + wi I
dt
j =1
(
)
4. 4
where the last term of the right side includes an external input I. However, the wi
parameter is ≠ 0 only for those genes (represented by nodes) biologically able to
receive the signaling transduced external input, in other words, only the nodes
representing the TF genes receive the external signal. It makes no biological sense
that all the nodes receive an input from the very beginning of the simulation runs,
because it can easily impose a serious bias to the resulting network.
Moreover, the input here generated has to be the same for the selected 4 input nodes
(JUNB, FOS, EST1 and EGR1) because no additional information exists. The only
other possibility to change the input that I consider could be changing the intensity of
the input that every one of the so-called input-nodes could receive. Therefore, the
range for the wi parameter ranges [-1,1] meaning that some genes are activated with
different strength by the input and even some others could be (which not necessarily
is the case) inhibited by the input. Changing the shape of every input is again a strong
assumption that drives easily the results towards a desired bias. As mentioned before ,
this is especially true if a larger time delay is chosen, as for instance 2h., because then
for the initial dynamics nodes are controlled mostly by the input. Therefore, here the
general input is modeled by an exponential function giving a dynamics as depicted in
figure 4.25 and as mentioned the wi parameter was optimized to be ≠ 0 only for the
ETS1, JUNB, FOS and EGR1 genes. Notice that there is a basal input along the entire
simulation runs.
Signal strength
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
121
2
1.5
1
0.5
0
1
35 69 103 137 171 205 239
Time (minn)
Figure 4.25 Fraction of the dynamics of the
simulated external input. This simulated
input represents a constant basal stimulus of
0.15 arbitrary units until the end of the
optimization lapse of 480 units’, equivalent
to 8 hours.
.
However, control data fittings were performed with no input at all, and the same input
without a basal signal. The rest of the parameters were chosen to be the same as those
for the Yeast and repressilator data sets. MSE results of the 50 data fitting runs per
each experiment, with and without input, are depicted in the Histograms of Figure
4.26 a and b respectively
Two important observations need to be mentioned from these histograms. The first is
that no significantly different distributions were found among the two of them (t-test
at p= 0.05). The second is that the fitting was performed to very low MSE (high
fitness) ranges in all cases (no outliers). This results indicates that the data is very
little restrictive and the fitting of the data has been very easily achieved.
Unfortunately, this could indicate that the number of different solutions fitting the
data equally well would be very broad.
15
a)
10
b)
0
Bin
5
0
0.
12
0.
16
0.
20
0.
25
0.
29
0.
34
0.
38
M
or
e
5
Frequency
10
0.
12
0.
17
0.
21
0.
25
0.
30
0.
34
0.
38
M
or
e
Frequency
15
Bin
Figure 4. 26 Histograms of the fitting runs of the TDRNN model a) using an external input and
b) without external input
122
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Clustering and robust parameters identification
After clustering and performing the robust parameter identification as described in
Methods, only poorly connected networks were found for both optimization groups,
with and without input.
Therefore, the z-score threshold was lowered to 1 unit in other to impose a less
restrictive criterion for the identification of robust parameters in the clusters. The
resulting five networks encountered for the input bounded group are depicted in
Figure 4.27.
a)
b)
c)
d)
e)
Figure 4. 27 Resulting inferred networks of the five clusters of the fitting runs: a, b and e are discarded
because of lack of biological meaning; network c and d represent the potential solutions of the RE
selected data.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
123
The results depicted in figure 4.27 show that despite the fitting of the data were
performed very well, no common pattern could be found as in the previous work (see
Yeast in section 4.2). Notice that the index were not included since no significant
difference among them was found (0.22, 0.20, 0.25, 0.23, 0.21 for the a, b, c, d and e
networks). A similar situation was obtained for the case of the no input optimized
group. Hence, it is not possible to elaborate serious conclusions about the different
networks obtained.
However, the networks c and d exhibit a more realistic biologically founded scenario.
Usually, these two networks would be the only two considered for further analysis,
discarding the other three. Hence, these two networks would be an intermediary step
on the iterative process of RE of GRN. One would extract some connectivity
hypothesis of them and test them by simpler and faster experiment in the wet
laboratory.
On the other hand, additional efforts to improve the identification of a common
network pattern or robust parameters were performed. Analogous to the repressilator
case section 4.1.2, the further dynamics of every optimization was simulated for every
model by the equivalent of three times the optimized interval meaning 24h (1440
min). As expected, all the models of the no input run arrived to a steady state defined
by their last activity states. Therefore, all the encountered models remain activated.
This could be argued as the desired long term activity behavior found in migration.
However, due to the optimization scheme, I consider this result just as an artifact from
the indeterminacy of the system. Moreover, the models optimized including an
external input behaved interestingly with a robust similar behavior, even though the
prolonged simulations do not include any input. Both cases are exemplified by one
randomly selected model of every group, depicted in figure 4.28.
124
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
a)
Expression
10
8
6
4
2
0
-2
1
481
Time (min)
961
1441
b)
Expression
10
8
6
4
2
0
-2
1
481
961
1441
Time (min)
Figure 4. 28 dynamics stability analysis of two solutions of the
simulated data; a) the simulation of the model using the
external input until 480 min, b) simulation of the model without
the external input
Notice that after withdrawing the input at the time of 8h (480 min), the model
depicted in Figure 4.28 presents only a small perturbation, but it recovers by itself in
the same steady state. This stable steady state was confirmed for longer intervals than
50h. Therefore, in my case, the fitting of this data drives the model to a stable steady
state. However, despite this robust behaviors are logical and expected, it was still not
possible to find a differential cluster in the solutions space of every group.
Definitively the modeling and process of reverse engineering GRN is possible and
robust as previously shown in sections 4.1 and 4.2, but definitely it also has its
limitations. As previously mentioned, the more important gain from this section is to
notice the improvement opportunities for future works. For instance, for the secondly
mentioned issue of measuring over the mixture of cells, an experimental improvement
could be the use of the Flow cell Cytometry technique (Gray, et al., 2007). Then,
quantitative information could be obtained about the proportion of every cell into the
mixture. Then a numerical algorithm could be applied and signal deconvolution could
be applied (Fellenberg, et al., 2001) to split the measured mRNA intensity signal
according to the proportional contribution of every cellular line.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
125
Additionally, to extend the cell system dynamics sampling would be another easy
improvement for any data set like the here analyzed. As previously mentioned, having
just one measurement after one hour after the supplied stimuli, it opens the possibility
for behaviors like the showed by the PTGS2 gen. Notice that certainly in the network
depicted in figure 4.27e this protein encoding gene appears at the top of an activation
cascade. However, this simple improvement could be achieved only if the
experimental setup is of a reproducible nature.
126
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
127
5. Discussion
In general, the reverse engineering of gene regulatory networks serves two purposes.
It is concerned in the development of models with high predictive power to perform
in silico experiments, as for instance theoretical knock outs. However, the main task
of RE of GRN is to use such models into a broader pipeline to establish a frame work
with the inference power to generate new knowledge about the biological networks
under study. Therefore, this is a hot topic into modern Biological Sciences and new
models and frameworks are intensively developed during the last ten years. However,
since this is a multidisciplinary area, in fact it requires a deep knowledge on different
areas such as molecular biology, mathematics, computer science, biochemistry etc.
This situation has been source of limitations and even several misconceptions in the
development of such models and frame works in the area. In this section, I will try to
summarize the more important results obtained in this thesis, pointing out the critical
steps for the area, and finally suggest some areas of improvement according to the
experience gained along this work.
128
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
5.1 Model choice and data driven experiments
As explained in the Biological context, traditional enzymatic models have no chance
to deal with the complexity of gene regulatory networks because the available data is
by far not enough in quantity or in quality to perform such calculations. Instead,
pragmatical models, whether black or grey could be useful ones, depending on
particular research objectives.
In the introduction and along this thesis it was demonstrated that, the new
technologically generated data are playing a double role in the area; on one hand they
are the motivation for such an integrative theoretical work, but at the same time they
are the bottleneck to obtain more accurate models and inference frame works.
Therefore, in the last years the international community has started to deal with this
situation and started to perform the so-called data-driven experiments. Data driven
experiments are tedious and expensive from an experimental biologist point of view.
They require, for instance, tightly sampled time-series for every stimulus-response
experiment and with they respective (statistically significant) replicates. However,
this data driven experiments are absolutely necessary to any serious effort in modern
Systems Biology.
Sooner or later the data limitation will be solved and large amounts of data will need
to be integrated into a new nascent biological epistemology. Then, I will come back to
the problem of the choice of the adequate model, according to every particular
research goal. In this sense, there are good and very interesting efforts to avoid
systematically the misconceptions and misunderstandings that the bias from having
different professional formations imposes over the choice of a model or a framework.
In this context, I propose with this work, a semiautomatic (globally optimized) fast
generation of models to cover, at a topological description level, two biological
networks: Signal transduction and gene regulatory networks, through a general time
delayed recurrent neural network. Additionally, I am proposing the innovative
parameter clustering technique of the models solutions, to improve the actual
inference power of different frameworks. However, there are still many issues to
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
129
fulfill in order to improve the reverse engineering of actual frame works. These issues
are further discussed into next sections of this general discussion chapter.
5.2 Data selection
Following the order of the proposed workflow (see figure 2.1), the first critical step in
the RE area is the data selection to reduce the dimensionality of the system to be
engineered. Besides the different techniques described in the sections of related works
and methods, a central problem needs to be solved: the so called data orthogonality.
Ideally the selected genes or proteins (nodes, from a modeling perspective), data
should be autonomous (or orthogonal) for the cellular function under study, with
respect to the rest of the cellular components. These selected nodes should include
sufficient information that none of the gene or proteins represented have missing
information to explain its activation or inhibition, as it occurs in the reverse
engineered network proposed by Li et al. where five auto inhibitory feedback loops
needs to be added in order to make the module functional.
The gene regulatory goal network of the yeast cell cycle used in this study is a
conceptual approximation based on experimental evidence. Here, I used just a single
gene expression time series of unknown quantity of noise for reverse engineering.
Hence, it would have been rather unlikely that all inferred connections would be
correct. In this light, I consider that finding the 44% of the total correlations (15/34),
or the 54% (7/13) of the indispensable weighted directed edges is a very good result.
Obviously, I do not suggest testing experimentally for the false positives I
encountered. Instead I suggest giving them a reading in the GED context, taking into
account meaningful shortcuts.
In the section of the repressilator case study 4.1.4, the incompleteness of information
in respect to the number of optimized nodes, I included one or two “blind” nodes into
the model to fulfill the missing nodes. As this is valid for a benchmark to analyze the
inference power of the models, it could not be the case for a normal pipeline as in the
keratinocytes case study here analyzed. Usually, one has no precise information of the
130
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
percentage of missing nodes. The solution in this case was to use my bottom up
proposal, starting from a core of nodes which was selected by using experts’
knowledge and making it grow whenever more nodes needed to be added in the light
of biological interest. However, the used approach in a related keratinocytes case
study (Busch, et al., 2008) differs in this by the use of an expression strength ranking
of the nodes (in this case genes). Since the expression range of the next added nodes
is considerably smaller, for sure, they will not change the initial engineered topology,
this solution imposes a strong bias to the original selected core based on the function
one is interested to find.
I rather proposed to work under a bottom up approach, started with the minimum
nodes information of biological interest, selected without a bias, by for instance the
GeneSet enrichment platform (see Methods). Additionally, once an initial nodes core
is selected and reverse engineered, one should make it grow by adding “blind nodes”
for a < 25% range of the total initial nodes core and reengineer it until one can define
a common topological pattern. Then, one can fix the encountered robust parameters;
add new nodes data and restart the reverse engineering process. For sure, this is a lot
of work to be performed, but imposes no bias towards a particular result. Therefore,
the time consumption could be compensated by the knowledge one can obtain into
this supervised iterative process.
5.3 Data interpolation, implications
Surprisingly, the TDRNN model is very robust against noise to perform the RE task.
However, the different interpolation techniques are playing an important role to
represent noisy data. My results are in agreement with : “You should not fit data with
a parametric model after smoothing, because the act of smoothing invalidates the
assumption that the errors are normally distributed” (Draper and Smith, 1998).
Therefore, for higher quantities of noise (> 20%) the linear interpolation performs
better on the synthetic data set. Consistently, the results from the experiments in
section 3.2 suggest that interpolation plays an important role to facilitate the
approximation of the data. Here, the MSE of the TDRNN using linear interpolation is
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
131
twice as low as when the same model is using another interpolation approaches.
Therefore, the relationship between these MSE and the GED results suggests that this
large difference is playing a significant role to the entire RE process. Hence, the
results suggest the use of linear interpolation unless the percentage of noise is low.
The superior reverse engineering performance of my TDRNN model compared to the
CTRNN model on the synthetic network was obtained under different circumstances
such as incomplete or noisy data. This achievement was obtained despite the fact that
this data set does not have a different time delay between genes and is highly
symmetric, which supposes an advantage for the CTRNN model because it already
has synchronized node responses.
5.4 Data fitting and inference power relationship
The next critical step into any RE framework is, to define the optimal data fitting
scheme used in every particular problem. Notice, this includes any additionally
considered information as the sparsely mixed optimization function here introduced.
Since the MSE is not a normalized error measure (Battaglia, 1996) but depends on the
range of the data, it is not possible to determine a MSE threshold on different data sets
to distinguish good from bad data fitting. Therefore, we see for the shortest periods
[0.2, 0.33] of the 4.1.2 repressilator case study, that the less information to measure
the error, the lower is the MSE inter-group limit obtained (see lower whiskers from
boxplots in figure 4.7. However, even though these optimizations with the lower MSE
have a direct correlation with the errors (see figure 4.9), the provided information is
still too few to restrict the solution space, and even several unrelated solutions are
found (see figure 4.10, upper panels). In other words, for these low information
conditions the GA get trapped in local maxima with fitness levels as high as the
correct solution.
132
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Moreover, I have realized that fitness alone is a poor indicator with variable
correlation to the inference power (RE task) of a model. This is especially true for the
high fitness (low MSE) [0.5, 0.75]period regions in section 4.1.2 and what should be
seen as enough data [1, 3] period regions on the 4.1.2 experiment. Additionally,
notice that the easy-to-fit zone does not correlate with the anti-correlation zone among
1 to 1.5 periods. Indeed, correlation depends on both: data structure and relative
fitness values (as previously explained, the MSE is not an absolute error
measurement). This is complex at a first glance, but it is easier to understand through
some examples from the 4.1.2 experiments. Here, in figure 5.1 are plotted the
scatterplots of the individual-errors vs. the fitness expressed by the logarithm of the
MSE, for the smallest optimized interval (0.2 periods) of the four models.
5
15
TDRNN-NP
5
-5
0
Log(MSE)
CTRNN-P
10
5
0
-10
-5
0
Log(MSE)
10
0
-10
15
Errors
10
0
-10
Errors
CTRNN-NP
15
Errors
Errors
15
-5
0
Log(MSE)
TDRNN-P
10
5
0
-10
-5
0
Log(MSE)
5. 1 scatterplots of the relationship between data fitting (log
MSE) and inference power (errors); for these fitness values, the
relationship follows a quadratic function.
As it can be seen in figure 5.1, the correlation in all these cases does not follow a
linear tendency but appears like a quadratic regression function. This is what is
expected for this insufficient information data structure. The interpretation is, that for
regions with poor fitting values (right side of the vertex of every parabola, MSE ≥ 1)
the models are facing problems to fit the data, therefore there are several individualerrors and therefore, one faces a strong positive correlation.
For imperfect data ( as for only 0.2 period), it occurs that fitting this imperfect data
structure whit high accuracy (left side of the parabolas, negative log (MSE)) is
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
133
associated to a wrong network topology, a bad generalization which is an analogous
of over-fitting. This occurs because the optimization process of the models easily
finds different solutions than the expected one which could better fit that imperfect
data structure. Therefore, the number of individual-errors increases according to the
lowest mean value of the quadratic regression at the vertex of the parabolas. This
parabolic behavior is exactly what is expected when over-fitting problems arise.
Obviously one can ask how it is possible then that on the figure 4.9 all the correlations
for this period data show a clear positive correlation. The answer is in the boxplots of
figure 4.7, this data structures (0.2 periods) have the largest distributions of MSE,
ranging from very small MSE with high fitness, shown by the lower whiskers, to
large MSE with few outliers, meaning covering from over fitting regions to regions
with poor fitting. However, despite having few outliers the distributions are not
normal and therefore, the Spearman correlation were used. Contrary to the Pearson
correlation which is based on linear correspondence, Spearman is based on ranked
data and therefore a relative strong correlation could still be observed for these data
regions. As previously mentioned, one should not base important conclusions just on
correlations without analyzing the scatterplots. However, here there are no
contradictions but an answer to the complex behavior of the correlation.
In principle, one should expect that the vertices of this parabolas moves towards the
origin of the graph (as depicted in figure 5.2 by a dashed arrow) with the increment of
information (longer optimized period regions). Then, there should be only a positive
correlation or no correlation when the fitness is around the vertex (as depicted by the
box around the vertex in figure 5.2). In the case of the easy–to-fit-regions of figure
4.7 occurs that all the fitting runs have a good fitness (low MSE) without over-fitting
and few fitting failure. Therefore, they correlation is around the vertex of this
parabolic correlation regression, meaning close to cero or no Spearman correlation.
134
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
5. 2 Proposed parabolic behavior of the inference power and
fitness relationship
Notice that a second issue for such kind of correlations is, that the individual-errors on
the y-axis are measured in a discrete scale while the fitness (MSE) has a continuous
scale. Hence, for small variations of the MSE around the vertex, there are no
variations of individual-errors which means no correlation. This explains why a
quadratic correlation regression would give the same results. But it is more important
to take into account the fitness (MSE) region where the optimization runs are. These
last two arguments partially explain the third fitness (MSE) optimization region,
Since the fitness alone is a poor indicator for the RE task, there should be other
criteria to optimize the models additionally, to the fitting of the data. For instance, one
should have an idea of the noise or incompleteness of the data in order to use the
“early stopping” criteria according to some reference. This reference is usually not
available, but could be or should be part of a routinely experimental setup. It could be
established by the use of any previous knowledge of the network to be engineered.
Another possibility is to perform the optimizations letting the model to explore for
different possible solutions (or fitness regions), then to utilize the clustering technique
to identify these different solutions and to evaluate them with the index here
suggested.
Regarding the network sparsity, which was thought as another criteria to optimize the
models, I found that the here introduced adaptive pruning function decreases the
connectivity of the resultant networks, while no considerable alteration of their fitness
occurs.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
135
While the pruning function helps to split the different network solutions and
consequently the clusters, I have found no evidence that pruning alone improves the
RE task. Now it is clear that network sparsity is not a decisive criteria for the RE task.
Instead, the connectivity characteristics as small world (Potapov, et al., 2005) or scale
free (Balaji, et al., 2006; Chen, et al., 2008; Iguchi, et al., 2007; Kauffman, 2004;
Wildenhain and Crampin, 2006; Zhou, 2005) networks could be taken into account .
However, in my case the size of the studied networks was not suitable for such an
analysis.
5.5 Reverse engineering framework, improving the robust
parameter selection
An effective improvement to the reverse engineering is the clustering of different
model solutions based on the identification of robust parameters. This clustering
approach is innovative since it is applied directly to the parameter space and not to the
dynamics of the models. This difference is exemplified in section 4.1.4. The use of
Lyapunov exponents to restrict solutions to those with stable and similar dynamics
could not find the two valid repressilators architectures, but the clustering applied to
the parameters solution space does it. Additionally, the same clustering approach
increased the number of correct edges (see. section 4.2.3), while not affecting the
number of false positives.
Moreover, I found the cluster index to be useful to distinguish between possible
solutions. Instead of a unique solution this strategy will usually narrow down putative
solutions to few possibilities to be validated experimentally.
136
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
137
6. Conclusions
Reverse engineering of gene regulatory networks is an iterative process between
experiments and modeling. In this process, the correct representation of the original
system is a critical step. In this sense the TDRNN has been shown to constitute an
improved approach as compared to existing CTRNN or DBN models. Additionally,
the clustering of the reverse engineering solutions provides a novel method to identify
robust parameters within the dynamic recurrent neural networks. Altogether, I
presented a supervised learning framework that helps to provide novel insight into
dynamic systems properties of genetic regulatory networks both from a biological and
theoretical point of view.
If the only data source is of transcriptional nature, such as RT-PCR or microarray
data, only transcriptional networks can be inferred. Cellular information processing,
however, is a feedback entangled process between protein signaling and gene
regulation for which combined transcriptomic and proteomic data are needed. I
consider the TDRNN to be the ideal model to incorporate both types of data. Due to
its incorporation of time delays that can potentially range from seconds to hours, it
can naturally incorporate fast responses occurring in signal transductions and slow
responses from the genetic regulatory network as we have demonstrated for the yeast
cell cycle data set.
138
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
139
7. Bibliography
Akutsu, T., Miyano, S. and Kuhara, S. (2000) Algorithms for identifying Boolean
networks and related biological networks based on matrix multiplication and
fingerprint function, J Comput Biol, 7, 331--343.
Albertini, F. and Sontag, E. (1993) Uniqueness of weights for recurrent nets,
Proceedings of the International Symposium Math. Theory of Networks Syst., II, 599602.
Arkin, A., Ross, J. and McAdams, H.H. (1998) Stochastic kinetic analysis of
developmental pathway bifurcation in phage lambda-infected Escherichia coli cells,
Genetics, 149, 1633-1648.
Auboeuf, D., Batsche, E., Dutertre, M., Muchardt, C. and O'Malley, B.W. (2007)
Coregulators: transducing signal from transcription to alternative splicing, Trends
Endocrinol Metab, 18, 122-129.
Bar-Joseph, Z. (2004) Analyzing time series gene expression data, Bioinformatics, 20,
2493-2503.
Bar-Joseph, Z., Farkash, S., Gifford, D.K. and Simon, I.R., R. (2004) Deconvolving
cell cycle expression data with complementary information, Bioinformatics, 20 Suppl
1, I23-I30.
Battaglia, G.J. (1996) MSE, AMP Journal of Technology, 5, 31-36.
Bay, S.D., Shrager, J. and Pohorille, A.L., P. (2002) Revising regulatory networks:
from expression data to linear causal models., J Biomed Inform, 35, 289-297.
Bebis, G.G., Michael. Kaspalris, Takis.
(1996) Coupling weight elimination and
genetic algorithms. Neural Networks, 1996., IEEE International Conference on.
Washington, DC, USA, 1115-1120.
140
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Beer, R.D. (1995) On the dynamics of small continuous-time recurrent neural
networks, Adaptive Behavior, 3(4), 469-509.
Beer, R.D. (2006) Parameter space structure of continuous-time recurrent neural
networks, Neural Comput, 18, 3009-3051.
Ben-Dov, C., Hartmann, B., Lundgren, J. and Valcarcel, J. (2008) Genome-wide
analysis of alternative pre-mRNA splicing, J Biol Chem, 283, 1229-1233.
Blackwood, E.M. and Kadonaga, J.T. (1998) Going the distance: a current view of
enhancer action, Science, 281, 60--63.
Blanco, A. and Delgado, M.P., M. C. (2001) A real-coded genetic algorithm for
training recurrent neural networks, Neural Netw, 14, 93-105.
Boutros, M. and Ahringer, J. (2008) The art and design of genetic screens: RNA
interference, Nat Rev Genet, 9, 554-566.
Busch, H., Camacho-Trullio, D., Rogon, Z., Breuhahn, K., Angel, P., Eils, R. and
Szabowski, A. (2008) Gene network dynamics controlling keratinocyte migration,
Mol Syst Biol, 4, 199.
Cao, Y. and Gillespie, D.T.P., L. R. (2006) Efficient step size selection for the tauleaping simulation method, J Chem Phys, 124, 44109.
Cremer, T. and Cremer, C. (2001) Chromosome territories, nuclear architecture and
gene regulation in mammalian cells, Nat Rev Genet, 2, 292--301.
Chen, K. and Rajewsky, N. (2007) The evolution of gene regulation by transcription
factors and microRNAs, Nat Rev Genet, 8, 93--103.
Chen, T., He, H.L. and Church, G.M. (1999) Modeling gene expression with
differential equations, Pac Symp Biocomput, 29--40.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
141
D'Haeseleer, P., Liang, S. and Somogyi, R. (2000) Genetic network inference: from
co-expression clustering to reverse engineering, Bioinformatics, 16, 707-726.
D'Haeseleer, P., Wen, X., Fuhrman, S. and Somogyi, R. (1999) Linear modeling of
mRNA expression levels during CNS development and injury, Pac Symp Biocomput,
41-52.
D´Haeseleer, P. (2000) Reconstructing gene networks from large scale gene
expression data, Ph.D. Thesis.
de Jong, H. (2002) Modeling and simulation of genetic regulatory systems: a
literature review., J Comput Biol, 9, 67-103.
Downes, S.M. (2004) Alternative splicing, the gene concept, and evolution, Hist
Philos Life Sci, 26, 91-104; discussion 123-109.
Draper, N.R. and Smith, H. (1998) Applied Regression Analysis. John Wiley \& Sons,
New York.
Edwards, R. and Glass, L. (2000) Combinatorial explosion in model gene networks,
Chaos, 10, 691--704.
Elowitz, M.B.L., S. (2000) A synthetic oscillatory network of transcriptional
regulators, Nature, 403, 335-338.
Fellenberg, K., Hauser, N.C., Brors, B., Neutzner, A., Hoheisel, J.D. and Vingron, M.
(2001) Correspondence analysis applied to microarray data, Proc Natl Acad Sci U S
A, 98, 10781-10786.
Feyzi, E., Sundheim, O., Westbye, M.P., Aas, P.A., Vagbo, C.B., Otterlei, M.,
Slupphaug, G. and Krokan, H.E. (2007) RNA base damage and repair, Curr Pharm
Biotechnol, 8, 326-331.
Florin, L., Hummerich, L., Dittrich, B.T., Kokocinski, F., Wrobel, G., Gack, S.,
Schorpp-Kistner, M., Werner, S., Hahn, M., Lichter, P. and Szabowski, A.A., P.
142
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
(2004) Identification of novel AP-1 target genes in fibroblasts regulated during
cutaneous wound healing, Oncogene, 23, 7005-7017.
Friedman, N., Linial, M., Nachman, I. and Pe'er, D. (2000) Using Bayesian networks
to analyze expression data, J Comput Biol, 7, 601--620.
Funahashi, K. (1989) On the Approximate Realization of Continuous Mapping By
Neural Networks, Neural Networks, 2, 183-192.
Gillespie, D.T. (1977) Exact stochastic simulation of coupled chemical reactions, J.
Phys. Chem., 81, 2340-2361.
Gillespie, D.T. (1992) A rigorous derivation of the chemical master
equation, Physica A, 188, 404-425.
Gray, A.C., McLeod, J.D. and Clothier, R.H. (2007) A review of in vitro modelling
approaches to the identification and modulation of squamous metaplasia in the human
tracheobronchial epithelium, Altern Lab Anim, 35, 493-504.
Guet, C.C., Elowitz, M.B. and Hsing, W.L., S. (2002) Combinatorial synthesis of
genetic networks, Science, 296, 1466-1470.
Hancock, E.R. (2005) Graph Edit Distance from Spectral Seriation, IEEE Trans.
Pattern Anal. Mach. Intell., 27, 365--378.
Hasty, J., Pradines, J., Dolnik, M. and Collins, J.J. (2000) Noise-based switches and
amplifiers for gene expression, Proc Natl Acad Sci U S A, 97, 2075--2080.
He, L.H., G. J. (2004) MicroRNAs: small RNAs with a big role in gene regulation.,
Nat Rev Genet, 5, 522-531.
Hill, T. and Lewicki, P. (2006) Statistics : methods and applications : a
comprehensive reference for science, industry, and data mining. Tulsa, Oklahoma.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
143
Hoops, S., Sahle, S., Gauges, R., Lee, C., Pahle, J., Simus, N., Singhal, M., Xu, L.,
Mendes, P. and Kummer, U. (2006) COPASI--a COmplex PAthway SImulator,
Bioinformatics, 22, 3067-3074.
Hu, X., Maglia, A. and Wunsch, D. (2005) A general recurrent neural network
approach to model genetic regulatory networks, Conf Proc IEEE Eng Med Biol Soc,
5, 4735-4738.
Hu, X., Maglia, A., Wunsch, D. (2005) A general recurrent neural network
approach to model genetic regulatory networks. , In Conf Proc IEEE Eng Med Biol
Soc, 5, 4735 – 4738
Huber, W., von Heydebreck, A., Sultmann, H., Poustka, A. and Vingron, M. (2002)
Variance stabilization applied to microarray data calibration and to the quantification
of differential expression, Bioinformatics, 18 Suppl 1, S96-104.
Itzkovitz, S., Tlusty, T. and Alon, U. (2006) Coding limits on the number of
transcription factors, BMC Genomics, 7, 239.
Juliano, R., Alam, M.R., Dixit, V. and Kang, H. (2008) Mechanisms and strategies for
effective delivery of antisense and siRNA oligonucleotides, Nucleic Acids Res, 36,
4158-4171.
Kastner, J., Solomon, J. and Fraser, S. (2002) Modeling a hox gene network in silico
using a stochastic simulation algorithm, Dev Biol, 246, 122--131.
Kauffman, S. (1971) Gene regulation networks: a theory for their global structure and
behaviors, Curr Top Dev Biol, 6, 145-182.
Kauffman, S., Peterson, C. and Samuelsson, B.T., C. (2004) Genetic networks with
canalyzing Boolean rules are always stable, Proc Natl Acad Sci U S A, 101, 1710217107.
144
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Keasling, J.D., Kuo, H. and Vahanian, G. (1995) A Monte Carlo simulation of the
Escherichia coli cell cycle, J Theor Biol, 176, 411--430.
Kim, S.-S. (1998) Time-delay recurrent neural network for temporal
correlations and prediction, Neurocomputing 20 20, 253—263
Lanctôt, C., Cheutin, T., Cremer, M., Cavalli, G. and Cremer, T. (2007) Dynamic
genome architecture in the nuclear space: regulation of gene expression in three
dimensions, Nat Rev Genet, 8, 104--115.
Lewis, B.A. and Reinberg, D. (2003) The mediator coactivator complex: functional
and physical roles in transcriptional regulation, J Cell Sci, 116, 3667--3675.
Li, F., Long, T., Lu, Y. and Ouyang, Q.T., C. (2004) The yeast cell-cycle network is
robustly designed, Proc Natl Acad Sci U S A, 101, 4781-4786.
Liang, S., Fuhrman, S. and Somogyi, R. (1998) Reveal, a general reverse engineering
algorithm for inference of genetic network architectures, Pac Symp Biocomput, 18-29.
Liao, X. and Wang, J. (2003) Global dissipativity of continuous-time recurrent neural
networks with time delay, Phys Rev E Stat Nonlin Soft Matter Phys, 68, 016118.
Lipan, O.W., W. H. (2005) The use of oscillatory signals in the study of genetic
networks, Proc Natl Acad Sci U S A, 102, 7063-7068.
Ma, J. (2004) The capacity of time-delay recurrent neural
network for storing spatio-temporal sequences, Neurocomputing 62 19 – 37.
MacQueen, J.B. (1967) Some Methods for classification and Analysis of Multivariate
Observations, Proceedings of 5-th Berkeley Symposium on Mathematical Statistics
and Probability", Berkeley, University of California Press, 1, 281-297.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
145
Mathayomchan, B.B., R. D. (2002) Center-crossing recurrent neural networks for the
evolution of rhythmic behavior, Neural Comput, 14, 2043-2051.
McAdams, H.H. and Arkin, A. (1997) Stochastic mechanisms in gene expression,
Proc Natl Acad Sci U S A, 94, 814--819.
Misteli, T. (2001) Protein dynamics: implications for nuclear architecture and gene
expression, Science, 291, 843--847.
Mjolsness, E. and Sharp, D.H.R., J. (1991) A connectionist model of development., J
Theor Biol, 152, 429-453.
Murphy, K.a.M., S. (1999) Modelling Gene Expression Data using Dynamic
Bayesian Networks, Computer Science Division, University of California, Life
Science Division. Lawrence Berkeley National Laboratory.
Paddison, P.J. (2008) RNA interference in mammalian cell systems, Curr Top
Microbiol Immunol, 320, 1-19.
Pe'er, D., Regev, A., Elidan, G. and Friedman, N. (2001) Inferring subnetworks from
perturbed expression profiles, Bioinformatics, 17 Suppl 1, S215--S224.
Perkins, T.J., Jaeger, J. and Reinitz, J.G., L. (2006) Reverse engineering the gap gene
network of Drosophila melanogaster, PLoS Comput Biol, 2, e51.
Pilpel, Y. and Sudarsanam, P.C., G. M. (2001) Identifying regulatory networks by
combinatorial analysis of promoter elements, Nat Genet, 29, 153-159.
Potapov, A.P., Voss, N., Sasse, N. and Wingender, E. (2005) Topology of mammalian
transcription networks, Genome Inform, 16, 270--278.
Radde, N. and Kaderali, L. (2007) Bayesian Inference of Gene Regulatory Networks
Using Gene Expression Time Series Data. In. Springer Berlin / Heidelberg, 1-15.
146
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Rao, A., Iii, A.O.H., States, D.J. and Engel, J.D. (2007) Inferring time-varying
network topologies from gene expression data, EURASIP J Bioinform Syst Biol,
51947.
Reinitz, J.S., D. H. (1995) Mechanism of eve stripe formation, Mech Dev, 49, 133158.
Robles-Kelly, A. and Hancock, E.R. (2005) Graph edit distance from spectral
seriation, IEEE_J_PAMI, 27, 365--378.
Savageau, M.A. (1969) Biochemical systems analysis. II. The steady-state solutions
for an n-pool system using a power-law approximation, J Theor Biol, 25, 370-379.
Schreiber, M., Kolbus, A., Piu, F., Szabowski, A., Mohle-Steinlein, U., Tian, J.,
Karin, M., Angel, P. and Wagner, E.F. (1999) Control of cell cycle progression by cJun is p53 dependent, Genes Dev, 13, 607-619.
Shannon, P., Markiel, A., Ozier, O., Baliga, N.S., Wang, J.T., Ramage, D., Amin, N.,
Schwikowski, B. and Ideker, T. (2003) Cytoscape: a software environment for
integrated models of biomolecular interaction networks, Genome Res, 13, 2498-2504.
Slepoy, A., Thompson, A.P. and Plimpton, S.J. (2008) A constant-time kinetic Monte
Carlo algorithm for simulation of large biochemical reaction networks, J Chem Phys,
128, 205101.
Smolen, P. and Baxter, D.A.B., J. H. (2000) Mathematical modeling of gene
networks, Neuron, 26, 567-580.
Spellman, P.T., Sherlock, G., Zhang, M.Q., Iyer, V.R., Anders, K., Eisen, M.B.,
Brown, P.O., Botstein, D. and Futcher, B. (1998) Comprehensive identification of cell
cycle-regulated genes of the yeast Saccharomyces cerevisiae by microarray
hybridization, Mol Biol Cell, 9, 3273--3297.
Spellman, P.T., Sherlock, G., Zhang, M.Q., Iyer, V.R., Anders, K., Eisen, M.B.,
Brown, P.O. and Botstein, D.F., B. (1998) Comprehensive identification of cell cycle-
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
147
regulated genes of the yeast Saccharomyces cerevisiae by microarray hybridization,
Mol Biol Cell, 9, 3273-3297.
Stamm, S. (2002) Signals and their transduction pathways regulating alternative
splicing: a new dimension of the human genome, Hum Mol Genet, 11, 2409-2416.
Stoll, G. and Rougemont, J.N., F. (2006) Few crucial links assure checkpoint
efficiency in the yeast cell-cycle network, Bioinformatics, 22, 2539-2546.
Sturn, A., Quackenbush, J. and Trajanoski, Z. (2002) Genesis: cluster analysis of
microarray data, Bioinformatics, 18, 207-208.
Subramanian, A., Tamayo, P., Mootha, V.K., Mukherjee, S., Ebert, B.L., Gillette,
M.A., Paulovich, A., Pomeroy, S.L., Golub, T.R. and Lander, E.S.M., J. P. (2005)
Gene set enrichment analysis: a knowledge-based approach for interpreting genomewide expression profiles., Proc Natl Acad Sci U S A, 102, 15545-15550.
Swain, M., Hunniford, T., Dubitzky, W., Mandel, J. and Palfreyman, N. (2005)
Reverse-engineering gene-regulatory networks using evolutionary algorithms and grid
computing, J Clin Monit Comput, 19, 329-337.
van Nimwegen, E. (2003) Scaling laws in the functional content of genomes, Trends
Genet, 19, 479-484.
van Someren, E.P., Wessels, L.F. and Reinders, M.J. (2000) Linear modeling of
genetic networks from experimental data, Proc Int Conf Intell Syst Mol Biol, 8, 355-366.
van Someren, E.P., Wessels, L.F.A. and Backer, E.R., M. J. T. (2002) Genetic
network modeling, Pharmacogenomics, 3, 507-525.
Wahde, M.H., J. (2000) Coarse-grained reverse engineering of genetic regulatory
networks, Biosystems, 55, 129-136.
148
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
Weaver, D.C., Workman, C.T. and Stormo, G.D. (1999) Modeling regulatory
networks with weight matrices, Pac Symp Biocomput, 112--123.
West, A.G., Gaszner, M. and Felsenfeld, G. (2002) Insulators: many functions, many
mechanisms, Genes Dev, 16, 271--288.
Whitley, D. (1993) "A genetic algorithm tutorial," Tech. Rep. CS-93-103. Department
of Computer Science, Colorado State University, Fort Collins, CO 8052.
Wilson, C.L. and Miller, C.J. (2005) Simpleaffy: a BioConductor package for
Affymetrix Quality Control and data analysis, Bioinformatics, 21, 3683--3685.
Wu, C.C., Huang, H.C., Juan, H.F. and Chen, S.T. (2004) GeneNetwork: an
interactive tool for reconstruction of genetic networks using microarray data,
Bioinformatics, 20, 3691-3693.
Wuensche, A. (1998) Genomic regulation modeled as a network with basins of
attraction, Pac Symp Biocomput, 89--102.
Reverse engineering of genetic networks with
time delayed recurrent neural networks and clustering techniques
149
Erklärung
Ich, David Camachio Trujillo, habe diese Dissertation selbst verfasst und mich dabei
keiner anderen als der von mir ausdrücklich bezeichneten Quellen und Hilfen bedient.
Experimentelle Daten bzw. Materialien, die nicht von mir selbst erhoben bzw.
hergestellt wurden, habe ich besonders kenntlich gemacht.
Ich habe an keiner anderen Stelle ein Prüfungsverfahren beantragt und diese
Dissertation auch nicht anderweitig in dieser oder anderer Form bereits als
Prüfungsarbeit verwendet oder einer anderen Fakultät als Dissertation vorgelegt.
Teile der vorliegenden Arbeit wurden im Vorfeld in Absprache publiziert:
•
David Camacho-Trujillo, Hauke Busch, and Roland Eils (2008): Reverse
Engineering Gene Regulatory Networks with Time Delay Recurrent Neural
Networks. Accepted for publication at Bioinformatics
•
Busch, H., Camacho-Trullio, D., Rogon, Z., Breuhahn, K., Angel, P., Eils, R. and
Szabowski, A. (2008) Gene network dynamics controlling keratinocyte migration,
Mol Syst Biol, 4, 199.
•
Schnickmann,S., Camacho-Trujillo, D., Bissinger, M., Eils, R., Angel, P.,
Schirmacher1, P., Szabowski, A., Breuhahn, K. (2008): 1AP-1 controlled
hepatocyte growth factor (HGF) activation promotes keratinocyte migration via
CEACAM1 and uPA/uPAR, Accepted at Journal of Investigative Dermatology
•
Camacho, D., Busch, H., Eils, R., Angel, P., Sbawoski, A.. (2008) Means and
mehtods for diagnosing metastasizing potentials of tumor cells. Patent Pend.
Heidelberg, den
_________________________________
David Camacho Trujillo
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
advertisement
© 2021