Modelling Approaches for the Integration of Different
Omics and Database-Information for
Clostridium acetobutylicum
vorgelegt von
Dipl.-Ing.
Sebastian Curth
aus Berlin
von der Fakult¨at III −Prozesswissenschaften
der Technischen Universit¨at Berlin
zur Erlangung des akademischen Grades
Doktor der Ingenieurwissenschaften
−Dr.-Ing. −
genehmigte Dissertation
Promotionsausschuss:
Vorsitzender: Prof. Dr. Leif-Alexander Garbe
1. Gutachter: Prof. Dr. Peter Neubauer
2. Gutachter: Prof. Dr. Peter G¨otz
3. Gutachter: Prof. Dr. Reinhard Guthke
Tag der wissenschaftlichen Aussprache: 21. M¨arz 2014
Berlin 2014
D 83
To family
Acknowledgements
First of all, I want to thank Professor Peter G¨otz for giving me the great op-
portunity to work on the COSMIC2 project. I owe him the start of this work
by his innovative dynamic model from which all approaches in this thesis could
germ. I am very grateful for the moral support when I started working with
principal components to develop new ideas. Many thanks go as well to Peter
Neubauer, to whom I owe the possibility to develop this thesis in the Biotechnology
department of TU Berlin. I am also thankful to Reinhard Guthke who confidently
acknowledged to become referee to this thesis.
Discussion with Professor Große-Wiesmann on data analysis and engineering-
approaches always showed me different aspects and facets of the problem. They
did not reduce the number of questions, on the contrary, but they helped finding
pragmatic approaches a lot.
I want to thank Katy Wolstencroft, who introduced me to Taverna. Your enthusi-
astic support for workflow-creation allowed the development of core aspects of
this thesis.
Many thanks go to my colleagues from all the different projects at the Biotech-
nology department of Beuth-Hochschule. Many many hours were spent in front
of confusing graphs and diagrams and I was trying to make sense of them with
your support: Susanne Wickert, Julia Rosenl¨ocher, Katharina Tomschek, Susanne
Fischer, Sabrina Fischer, Kunigunde Stephani-Kosin, Tanja Westphalen.
COSMIC2 also consisted of a lot of experimental work. Many very fruitful
discussions were held with Dr. Stefan Junne to start experimental work with
Clostridia. Thank you Mirja Rothe, your excellent technical skills were vital for
the project and inspiring for my own work. Thank you as well my two bachelor
students, Marcel Mehl, who established fermentations, and Andriy Grygorenko
who standardised experimental proceedings and measurements, data acquisition
would not have been possible without you. Further thanks for technical support
in many practical questions go to Thorsten Jamrath, Barbara Ritsche, Nabeel
Fattohi and Harald Gerullis.
Thanks to the entire COSMIC2 consortium, who provided spores, protocols and
v
support in data acquisition.
This work was made possible through financial support by the BMBF project
0315782B (Verbundprojekt COSMIC2 - Systembiologie des Butanolproduzenten
Clostridium acetobutylicum: Eine neue Quelle f¨ur Treibstoffe und Grundchemi-
kalien; Teilprojekt B.)
Abstract
This work establishes methods to investigate the butanol formation of Clostridium
acetobutylicum. In particular, the generation of two types of models will be extens-
ively discussed. Therefor, the required formal basis for the model evaluation and
the information technological standards will be introduced, e.g. the construction
of a local database of clostridial annotation in KEGG.
The first model is a static pathway-model that provides the integration of tran-
scriptome data into a metabolic-network model for visualisation and analysis
purposes. It is proposed to use a novel rule from boolean logic for data integration
to facilitate visual access to characteristics of the metabolic network. As con-
sequence, the postulation of experimental hypotheses is facilitated: The possibility
of a 3-hydroxybutyrate dehydrogenase activity in C. acetobutylicum is illumin-
ated. The resulting priority list from annotation transfer contains functional and
regulatory aspects of the data and the databases and it hereby offers an optimal
starting point to initiate experimental work.
The second model is a dynamic model that is used to map metabolome and
transcriptome data from fermentation experiments together. Its unique structure
allows a number of new analyses - and shows new problems. Its simulation results
suggest that the pH-shift in C. acetobutylicum can be solely related to transcript
dynamics. Optimisation strategies on the transcript level and the parameter level
of the model will be implemented and their results discussed.
Finally, the principal component analysis will be used to optimise computation
times of such a model and from this, two novel methods will be derived: dynamic
aspects of transcriptome data will be alternated to construct regulatory similar
expression profiles with different amplitudes, and genes will be classified according
to their regulatory similarity in a novel clustering approach.
vii
Abstract
Diese Arbeit umfasst die Methoden-Erstellung zur Erforschung der Butanol-
Bildung von Clostridium acetobutylicum
¨
uber in silico Modelle. Zwei dieser
Modelle werden hier ausf
¨
uhrlich besprochen und die notwendige Basis aus ma-
thematischen Formalismen f
¨
ur die Evaluierung sowie informationstechnischen
Herangehensweisen eingef
¨
uhrt, wie z.B. die Etablierung einer lokalen Datenbank
der Clostridien-Annotation aus KEGG.
Das erste zu besprechende Modell ist ein statisches Pathway-Model, das es
erm
¨
oglicht Transkriptom-Daten unter Zuhilfenahme eines metabolischen Netz-
werkes darzustellen und zu analysieren. Insbesondere wird eine boolesche Logik
diskutiert, die die Daten-Integration vollzieht. Charakteristische Eigenschaften
des Netzwerkes werden so hervorgehoben und metabolische Zust
¨
ande visuell
zug
¨
anglich gemacht. Die Bildung experimenteller Hypothesen wird erleichtert:
Hier wird die M
¨
oglichkeit einer 3-Hydroxybutyrate dehydrogenase Aktivit
¨
at in C.
acetobutylicum n
¨
aher beleuchtet. Die resultierende Priorit
¨
atenliste des Annotations-
Transfers beinhaltet sowohl funktionale als auch regulatorische Informationen aus
Datenbanken und Experimenten und bietet somit einen optimalen Startpunkt,
experimentelle Forschung zu initiieren.
Das zweite zu besprechende Modell ist ein dynamisches Modell, das benutzt wird,
um Metabolom- und Transkriptom-Daten aus Fermentations-Experimenten zu
vereinigen. Seine besondere Struktur erm
¨
oglicht eine Vielzahl neuartiger Ana-
lysen - bringt aber ebenso neuartige Probleme mit sich. Die Resultate deuten
darauf hin, dass der pH-Shift in C. acetobutylicm allein von der Dynamik der
Transkriptom-Daten abh
¨
angt. Optimierungsstrategien auf Transkript-Ebene und
auf Parameter-Ebene des Modells werden implementiert und ihre Resultate disku-
tiert.
Schließlich wird
¨
uber Hauptkomponenten-Analyse sowohl eine Methode zur Opti-
mierung der Laufzeiten eines solchen Modells gegeben als auch zwei neue Methoden
geschlussfolgert: Eine, um die zeitliche Dynamik der Transkriptom-Daten geeignet
zu variieren ohne regulatorische Profile zu ver
¨
andern, die andere um Gene anhand
ihrer regulatorischen ¨
Ahnlichkeit zu klassifizieren.
ix
Contents
Contents i
List of Abbreviations and Symbols v
List of Figures xi
List of Tables xiii
1 Preliminaries 1
2 Introduction 5
2.1 Motivation of Research ......................... 6
2.1.1 Butanol Resource Management ................ 6
2.1.2 The Historical ABE Production Process ........... 7
2.2 Biological Facts ............................. 10
2.2.1 What is Clostridium acetobutylicum? ............ 10
2.2.2 Biochemical Pathways ..................... 10
2.2.3 Engineering a Butanol Production Strain .......... 14
2.2.4 Concurrent Designs ...................... 15
2.3 Butanol Fermentation ......................... 16
2.3.1 Production of Butanol ..................... 16
2.3.2 Counteracting the Effects of Butanol ............. 17
2.4 Published Data ............................. 20
2.4.1 Analysis Techniques for Different Omics ........... 20
2.4.2 Data Treated Within This Thesis ............... 21
2.5 Summary and Thesis Proposal .................... 24
2.6 Outline ................................. 25
3 Automated Network Model Creation 27
3.1 Database-Harvesting .......................... 28
3.1.1 Existing Pathway-Models ................... 28
i
ii CONTENTS
3.1.2 Local Solutions ......................... 29
3.2 Network Analysis of the Clostridial Reactome ............ 30
3.2.1 Graph Definition and Notations ................ 30
3.2.2 Softwares For Visualisation of Graphs ............ 31
3.2.3 Graph Characterisation .................... 32
3.2.4 Data-Driven Network Generation - Methodology ...... 33
3.2.5 Summary ............................ 36
3.3 Data-Driven Pathways ......................... 39
3.3.1 Derivation of the Boundary Parameter ............ 39
3.3.2 Augmentation Characterises Solventogenesis ........ 42
3.3.3 Visualisation of the Standard Batch Fermentation ..... 43
3.3.4 Visualisation of the Acetic Acid Pulse Experiment in Batch 46
3.3.5
Visualisation of the pH-Shift Experiment in Continuous
Culture ............................. 47
3.3.6 Conclusions ........................... 53
3.4 Identification of Missing Reactions .................. 56
3.4.1 Comparative Approach .................... 56
3.4.2 Comparison of B. subtilis and C. acetobutylicum . . . . . . 58
3.5
Annotation Transfer a Case-Study: 3-HBDH Activity in Clostridium
64
3.5.1 Annotation Transfer Methods ................. 64
3.5.2 Collection of Experimental Indications ............ 65
3.5.3 Collection of Database Information .............. 66
3.5.4 Hypotheses for Annotation Transfer ............. 67
3.5.5 Results ............................. 68
3.5.6 Critical Evaluation ....................... 73
3.6 Final Conclusions ............................ 77
4 Automated Dynamic Model Creation 79
4.1 Historical Perspective ......................... 80
4.2 Derivation of the Dynamical Model .................. 82
4.2.1 Derivation of the Model .................... 83
4.2.2 Formalism ............................ 86
4.2.3 Integration of Time-Dependent Data ............. 88
4.3 Model Implementation ......................... 90
4.4 Evaluation of Experiments ....................... 93
4.4.1 Computer, Softwares, Data, Algorithms ........... 93
4.4.2 Parameter Estimations and Validation ............ 96
4.4.3 Validation-Results .......................104
4.5 Sensitivity Analysis ...........................106
4.5.1 Local Sensitivity Analysis (LSA) ...............106
4.5.2 Global Sensitivity Analysis (GSA) ..............110
4.5.3 Summary ............................115
CONTENTS iii
4.6 Final Conclusions ............................116
5 Principal Component Analysis In Modelling 119
5.1 Introduction of PCA ..........................120
5.2 Data Compression in the Dynamical Model .............121
5.3 Optimisation Approach via PCA ...................123
5.4 Clustering from Principal Components ................129
5.4.1 Introduction ..........................129
5.4.2 Geometric Approach to Clustering ..............130
5.4.3 Application of Angular Traits .................132
5.5 Final Conclusions ............................135
Bibliography 141
A Dynamic Model Equations 165
B Scripts 167
B.1 Taverna Workflows ...........................167
B.2 Static Model Scripts ..........................169
B.2.1 Model Creation .........................169
B.2.2 Creation of a Comparison Database .............170
B.2.3 Creation of 3-HBDH Candidates ...............170
B.3 Dynamic Model Scripts ........................171
B.3.1 Converting The KEGG-Database Into The Standard Format171
B.3.2 Integrating Data Into The Standard Format .........171
B.3.3 Model Simulation ........................171
B.3.4 Sensitivity Analysis ......................172
B.4 Principal Component Analysis ....................173
B.4.1 Dynamic Features .......................173
B.4.2 Clustering ............................173
List of Abbreviations and
Symbols
∆Gnetwork diameter
γ
parameter indicating community structure from the dis-
tribution of node neighbours
Xcell density
bboundary for up-regulation or down-regulation
C
counter of genes that occur in subsets from annotation
methods
ci, coconcentration of intracellular or extracellular compound
cXbiomass concentration
Cecc/clo/rad/str/bet node centralities
e∈ E edges in the MMM, reactions
Fpump rate
G, Ga, H
Graphs of the MMM, standard, augmented, standard &
augmented
J(a, b) Jaccard index between objects aand b
j= 1..NJgene numbers
k11|2maximal conversion rate of acetate kinase
k1|3maximal conversion rate of thiolase
k1|9maximal conversion rate of ethanol dehydrogenase
v
vi LIST OF ABBREVIATIONS AND SYMBOLS
k1|11 maximal conversion rate of phosphotransacetylase
k23|17
maximal conversion rate of the CoA-transferase that uses
acetic acid
k3|4
maximal conversion rate of lumped reaction of 3-hydroxybutyry-
CoA dehyrogenase, butyryl-CoA dehydrogenase and 3-
hydroxybutyryl-CoA dehydrogenase
k4|10 maximal conversion rate of butanol dehydrogenase
k4|5maximal conversion rate of phosphotransbutyrylase
k5|6maximal conversion rate of butyrate kinase
k63|47
maximal conversion rate of the CoA-transferase that uses
butyric acid
kx|y
maximal conversion rate that uses substrate x and trans-
forms it into product y
mXbiomass weight
nP((a, b))
similarity of two genes a,b according to their Pfam-motifs
ni, noamount of intracellular or extracellular compound
Pfa, Pfunction that maps genes onto pfam-motifs
R2coefficient of determination
Rsink loss of metabolites by outflow
ri, roreaction rate of intracellular/ extracellular compound
rsxp reaction rate from substrate s to product p
Rct, Rfunction that maps genes onto reactions
sintegrated similarity score from multiple methods
v∈ V nodes in the MMM, compounds
VC, VR, VLcell volume, reactor volume, liquid volume
x∈ X,X0,XMi
all genes, genes without reaction annotation, genes as
candidates retrieved by Method Mi
x∈ Xb(s) genes being up-regulated
x∈ Xb(s) genes being down-regulated
vii
3-HBDH 3-hydroxybutyrate dehydrogenase
B. subtilis/BS Bacillus subtilis
ABE Acetone-Butanol-Ethanol
Ac acetate
Ac-CoA acetyl-CoA
Ac-P acetyl-phosphate
AcAc acetoacetate
AcAc-CoA acetoacetyl-CoA
ack acetate kinase
AcOn acetone
bcd butyryl-CoA dehydrogenase
BdhAB,AdhE alcohol dehydrogenases
BHBu-CoA 3-hydroxybutyryl-CoA
BL3D BioLayout Express 3D
BLAST Basic Local Alignment Search Tool
bn billion
BRENDA BRaunschweig ENzyme DAtabase
BSU KEGG gene identifiers for B. subtilis
Bu butyrate
Bu-CoA butyryl-CoA
Bu-P butyryl-phosphate
buk butyrate kinase
BuOH butanol
CA C. acetobutylicum
CAC/CAP
KEGG gene identifiers for C. acetobutylicum genes on
chromsome or plasmid
CNA CellNetAnalyser
viii LIST OF ABBREVIATIONS AND SYMBOLS
CPD: KEGG compound identifiers
crt crotonase
Crt-CoA crotonyl-CoA
ctfAB CoA-transferase A/B
DSMZ
Deutsche Sammlung von Mikroorganismen und Zellkul-
turen
E/N edges to nodes fraction
EtOH ethanol
FBA Flux Balance Analysis
G(s) graph at state s derived from up-regulation
H(s)
graph at state s derived from up-regulation and augment-
ation
hbd hydroxybutyrate dehydrogenase
KEGG Kyoto Encyclopedia of Genes and Genomes
LD50 lethal dose
MCL Markov Clustering
MetaCyc Encyclopedia of Metabolic Pathways
MJ Megajoule
MMM Metabolite-Metabolite-Mapping
OD optical density
OECD
Organisation for Economic Co-operation and Develop-
ment
ORF open reading frame
p.a. per annum
PCA Principal Component Analysis
PEP phosphoenolpyruvic acid
pSOL1 Megaplasmid of C. acetobutylicum
ix
PTS phosphotransferase system
RN: KEGG reaction identifiers
RVP Reid Vapour Pressur
SBGN Systems Biology Graphical Notation
SBML Systems Biology Markup Language
SED-ML Simulation Experiment Description Markup Language
SOAP Simple Object Access Protocol
thl thiolase
yEd Graph Editor by yWorks
List of Figures
2.1 Acid and Solvent Pathway ......................... 11
3.1 Determination of bOver Time ...................... 41
3.2 Determination of bFor All Transcriptome Experiments ......... 42
3.3 Graph Parameters ............................. 43
3.4 Static Model - Early Batch Experiment ................. 44
3.5 Static Model - Late Batch Experiment .................. 45
3.6 Static Model - Early Batch Pulse Experiment .............. 46
3.7 Static Model - Late Batch Pulse Experiment .............. 48
3.8 Static Model - Continuous Experiment, Acidogenesis .......... 49
3.9 Static Model Augmentation - Continuous Experiment, Acidogenesis . 50
3.10 Static Model - Continous Experiment, Solventogenesis ......... 51
3.11 Static Model Augmentation - Continous Experiment, Solventogenesis . 52
3.12 Reduced MMM of B. subtilis ....................... 60
3.13 Map of Genes From B. subtilis to C. acetobutylicum .......... 61
3.14 Distribution of Edge Weights ....................... 63
3.15 Alternative Models of Butyric Acid Production ............. 67
3.16 Co-Cluster Approach ............................ 71
3.17 Map of 3-HBDH Candidates ....................... 72
3.18 Pfam-motifs of BLASTP .......................... 75
4.1 Butanol-Production Model ........................ 83
4.2 SBTOOLBOX2 Standard Model ..................... 91
4.3 Parameter-Estimation Batch Culture - Concentrations ......... 96
4.4 Parameter-Estimation Batch Culture - Reactions ............ 97
4.5 Parameter-Estimation Continuous Culture - Concentrations ...... 98
4.6 Parameter-Estimation Continuous Culture - Reactions ......... 98
4.7 Parameter Uncertainty of the CM ....................101
4.8 Model Comparison to Literature - Promoter ...............102
4.9 Model Comparison to Literature - Deletions ...............103
xi
xii LIST OF FIGURES
4.10 LSA - Transacetylase & Kinase ......................107
4.11 LSA - Thiolase & CoA-Transferase ....................108
4.12 LSA - Dehydrogenase, Transbutyrylase ..................108
4.13 LSA - Butyrate Kinase ...........................109
4.14 LSA - Solvent Production .........................109
4.15 GSA - Transacetylase & Kinase ......................112
4.16 GSA - Thiolase & CoA-Transferase ....................113
4.17 GSA - Dehydrogenase, Transbutyrylase .................113
4.18 GSA - Butyrate Kinase ..........................113
4.19 GSA - Solvent Production .........................114
5.1 PCA for r4|10 Optimisation ........................124
5.2 Dynamic Features Optimisation - r4|10 ..................125
5.3 Transcript Optimization of r4|10 - Extracellular Metabolites ......126
5.4 Transcript Optimization of r4|10 - Intracellular Metabolites ......127
5.5 Dynamic Features Optimisation - r4|5..................128
5.6 Tiling Concepts ...............................132
5.7 PCA in Clustering .............................133
5.8 PC Representation of Data - Variances ..................134
5.9 Clustering Results .............................135
List of Tables
2.1 Industrial Clostridium Species ....................... 8
3.1 States For Network-Generation ...................... 39
3.2 Summary of Graphs ............................ 40
3.3 Statistics of the Comparison Knowledge-base .............. 59
3.4 Proof of Concept of the Comparison ................... 62
3.5 Results of BLASTP ............................ 69
3.6 3-HBDH Motifs ............................... 70
3.7 Candidate Occurrence ........................... 73
3.8 Candidate Ranking ............................. 73
4.1 Compound Overview ............................ 84
4.2 Model Reaction Kinetics .......................... 87
4.3 Softwares .................................. 93
4.4 Parameter-Estimation Results .......................100
5.1 Angular Traits Overview ..........................133
xiii
Chapter 1
Preliminaries
They both savoured the strange warm glow
of being much more ignorant than ordinary people,
who were ignorant of only ordinary things.
Terry Pratchett
This chapter introduces the two aspects in which this thesis in embedded, a
problem-centred project, SysMO-COSMIC2, and a systems biological approach
that integrates different types of data.
2CHAPTER 1. PRELIMINARIES
On Usefulness
Systems biological strategies for several organisms were funded by the transnational
initiative SysMO to enhance European wide collaborations, This initiative is
divided into several sub-projects, one of them is SysMO-COSMIC2. This project
is the starting point of this dissertation because it follows an engineering approach
on a well known, widely treated problem: renewable energy production and
chemical key compound generation. In the scope of dwindling crude oil reserves
this problem requires recapitulation with modern biological and information
technological techniques. Such a sustainable technology is under development by
investigating the fermentation of Clostridium acetobutylicum and optimizing its
productivity.
On Work Distribution in SysMO-COSMIC
COSMIC funding was sustained over two periods of three years each. While in
the first period, development of suitable standard operating procedures (SOPs),
development of cloning techniques and fermentations of the wildtype of Cl. aceto-
butylicum were in focus of research, the second period was used to emphasise on
mutant generation and mutant fermentation.
The development of a SOP for fermentations was necessary to allow comparison
of different experiments carried out at different sites. This procedure describes
the set up of a continuous chemostat culture that is shifted between two distinct
metabolic states by fermenting at two different pH values. The recorded responses
are thought to give major insights into the regulatory mechanisms of the organism.
This work was distributed along three modelling groups and five experimental
groups and included the generation of suitable mutation technology, the culturing
of mutant strains, the establishment of downstream protocols and the generation
of models to describe and design experiments.
On Data Deposit
Parallel to the proposition of experiments, a computer scientific issue is challenged
within all SysMO projects: Quantity and size of data require meaningful ways
of organization in a database, of annotation and of standardisation. Identically,
modelling approaches require description of the inherent model structure, graphical
representation of the model details, and simulation annotation. To tackle this
problem a sustainable platform called SysMO-SEEK was established to allow
exchange and future use of data and models. This is achieved by implementing
several standard formats.
3
On Standards
The System’s Biology Markup Language (SBML) [
Hucka et al., 2003
] aims at an
unified description of biological models. It is a XML-based format, distributed
as level 3, and it is widely acknowledged. Usually, model deposition into online
resources as e.g. Biomodels [
Li et al., 2010
] requires SBML format. Rapid brows-
ing through published models, downloading them and checking the published
results is one feature of this standard format. Reproducibility of published results
can be ensured using SED-ML, the Simulation Experiment Description Markup
Language [
Waltemath et al., 2011
]. Standardisation of graphical representations
recently started by the use of SBGN, the Systems Biology Graphical Notation
[
Klipp et al., 2007
]. Finally, integration of software tools into web-pages, like
JWS [
Snoep and Olivier, 2002
] in SysMO-SEEK allow also the experimentalist
to access and use models. Also the use of Taverna, which will be used later (3.1)
is only possible through this standardisation. The ongoing research on suitable
SBML features provides also a framework to standardise experiments. It will be
shown that the transferability of XML formats between different tools is an issue
however [
Alves et al., 2006
] and also SBML is not yet flexible enough to consider
all models of a certain type. Lately, efforts were successful in unifying standards
for pathway models in with a plug-in for Cytoscape [
Shannon et al., 2003
] which
is named BiNoM. It is a promising tool for integration of data and interoperability
of standards [Bonnet et al., 2013].
On Automation
The different approaches for data-analysis, data-curation and data-integration
easily outnumbers the data-creation effort [Palsson and Zengler, 2010]. Automa-
tion represents a possible tool to increase reproducibility of results and reduce
this effort. It will be shown that even knowledge discovery can be undertaken by
an automation approach [
Aksenov et al., 2005
]. Therefor, the scope of this work
is broaded. It is focussed on Cl. acetobutylicum research but it may be equally
used for any other organism.
On Storage of Results
Data and models presented in this thesis will be linked to my personal profile
∗
on
SEEK.
Reading of The Thesis
I encourage the electronic reading of this thesis, because the representation of
large networks is only poor in printed style. High zooms are supported by most
∗https://seek.sysmo-db.org/people/319
Chapter 2
Introduction
Here’s what I think the truth is:
We are all addicts of fossil fuels in a state of denial,
about to face cold turkey.
Kurt Vonnegut
This introduction is dedicated to give the background of this thesis, microbial
butanol production. First, it will give a historical motivation for the research
of the biological process that uses Clostridium acetobutylicum (2.1). Past and
current studies focussed on the biological and biochemical factors involved in
butanol synthesis (2.2). Also, a bouquet of process variants were researched to
increase productivity (2.3). Combining these information will help in focussing
and understanding several key experiments from literature. These acquired data
will play a dominant role in the following three chapters (2.4).
Finally the thesis proposal (2.5) and the thesis outline (2.6) will be given.
6CHAPTER 2. INTRODUCTION
2.1 Motivation of Research
This thesis’ core is the production of one chemical compound, butanol. From the
overview of its economical relevance (2.1.1), the necessity of alternative production
routes becomes obvious. Historically, a similar scenario was present when oil
refinery industry had not yet been developed. In these times, a biological process
had been used to generate acetone. It was named after its main products, acetone,
butanol and ethanol, the ABE-fermentation. This historical process was already
subjected to several optimisation approaches (2.1.2).
2.1.1 Butanol Resource Management
The Uses of Butanol
Butanol (CAS:71-36-3) serves as gasoline additive [
Duerre, 2007
] and as interme-
diate for the chemical production of acrylates, glycol ethers, resins, and various
esters. It further serves as solvent for various products, e.g. paints, gums, fats,
waxes, rubber, as a swelling agent and colour carrier in textile industry and as an
extraction agent for various drugs, antibiotics and hormones. It is also an additive
for the cosmetic and the cleaning industry [
Company, 2006
,
SE, 2008
]. In 2011,
the market volume of butanol is accounted to be 3 million tonnes, an increase
by 2
.
1 % compared to the previous year [
Tanya Rezler, 2012
]. The total market
price is estimated to rise from $5.9bn from 2011 to $9.2bn in 2015.
The Chemical Production of Butanol Relies on Oil Resources
The chemical process of butanol synthesis comprises the hydroformylation of
propylene with carbon monoxide to butyraldehyde in the presence of a rhodium-
based catalyst, which is followed by hydrogenation of the aldehyde to the al-
cohol [
Siegel and Himmele, 1980
]. In contrast to this oxosynthesis, the Reppe
synthesis directly produces butanol from propylene, it is however more expens-
ive [
Lee et al., 2008b
]. Since propylene is a product of the oil cracking process
[
Li et al., 2007
], butanol stands in direct relation with the availability of oil re-
sources.
New Opportunities Through Rising Oil Prices
The annual report of British Petroleum (BP) from 06/2012 [
Dudley, 2012b
] sum-
marises that the price for a barrel oil rose by 40 % from 2010 to 2011, which
makes the actual price $111.26 per barrel. This dramatic increase is not only the
result of dwindling crude oil resources, it can be also understood as a question for
a new process design of fuel production, that is covering both, world wide demand
and economical feasibility. Demand is increasing slower than supply (0.6 million
barrels per day vs 1.3 million barrels per day) which corresponds to growth by
2.1. MOTIVATION OF RESEARCH 7
0
.
3 % and 1
.
3 %, respectively. The largest increase in consumption was registered
in China, 505
.
000
barrels/day
. These numbers may sound enthusiastic, yet, pro-
jections from 01/2012 into the year 2030 estimate the increase of the global energy
consumption by 1
.
6 % p.a, leading to an increase by 39 % compared to nowadays
[
Dudley, 2012a
]. The major increase is due to non OECD-countries. Any estimate
that the consumption of raw materials may be regressive would be a fatal error
and new supply methods must be found. A second major issue is the import
dependency of crude oil for any country. Europe’s market is constantly and almost
entirely relying on the import of oil (94 % in 2030) and this dependency is even
increasing for gas (80 %). Industrial research consequently increases investments
into renewables, which are the fastest growing fuels by 8
.
2 % p.a. Butanol as
biofuel has a market value of $2.5 per gallon [
Pfromm et al., 2010
], while for the
chemical industry this price increases to $5-6 per gallon [
Doris de Guzman, 2011
].
Is Bio-Butanol a Competitor of Bio-Ethanol?
Bio-butanol production is considered an antagonistic product to bio-ethanol:
As fuel additive it has superior properties compared to ethanol. The Reid
vapour pressure (RVP) which is a measure of evaporative emissions is decreased
by a factor of 6 which makes butanol safer to handle. More importantly, its
improved hydrophobicity helps blending at higher concentrations with gasoline.
Engine modifications are not required. Ethanol can only be blended up to 85 %.
Hydrophobicity and decreased corrosiveness to metallic compounds of the pipelines
make butanol an ecologically safer compound regarding ground water. Last but
not least, the caloric value of butanol is just 10 % less than gasoline, and 50 %
higher than ethanol [
Brekke, 2007
,
Duerre, 2007
]. Environmental risk estimations
consider it as readily biodegradable under aerobic conditions. Acute toxicity in
water is reached at 0
.
5
g
L
. Importantly, it has low potential to accumulate in a
biological system. Studies showed that in animal systems the
LD50
ranged between
0
.
8
−
4
g/kg body mass
.In vivo hydrolysis of butanol occurs fast, 20
min
after
application of radioactive butyl acetate (30.2mg/kg of body weight) to rats, the
hydrolysis product butanol was not detectable anymore [Hernandez, 2004].
Nevertheless, it is argued that the current production process can not compete
with bio-ethanol production as long as feed-stock costs represent a major factor
of the production costs. A megajoule energy costs $0
.
07 for butanol and $0
.
03
for ethanol. The authors further argue that the current process robustness is not
sufficient for an industrial scale production [Pfromm et al., 2010].
2.1.2 The Historical ABE Production Process
This section summarises the very exhaustive review by Jones and Wood
[Jones and Woods, 1986] if not stated otherwise.
8CHAPTER 2. INTRODUCTION
Research Aimed at Acetone as Primary Product
Originating from a research project for rubber synthesis in Great Britain, in
the period from 1912-1914, Chaim Weizmann isolated a butanol and acetone
producing strain that was able to grow on potato starch and a broad range of other
polysaccharide substrates, as root crops, nitrogen-fixing legumes, cereal crops and,
more generally agricultural soil. The resulting patent became the initial point of
research for the following generations. Acetone as a base chemical for colloidal
production had economical priority. The historical process of acetone production
took calcium acetate and disrupted it into calcium oxide and acetone with the
help of heat. Since during the First World War imports of calcium acetate were
stopped, the British economy had to find alternative production routes, and the
several clostridial species were promising candidates (table 2.1), one of them is
nowadays in focus of research, Clostridium acetobutylicum.
Table 2.1: Clostridium species of interest to chemical industry in the DSMZ
database
Species DSMZ ID
C. acetobutylicum 792
C. saccharobutylicum 13864
C. butyricum 10702
C. pasteurianum 525
C. beijerinckii 791
C. tyrobutyricum 2637
Competition Inspired Optimization of Substrate Utilisation
The research focus has been the growth of Clostridium acetobutylicium on a
multitude of substrates: monosaccharides like lactose, polysaccharides like cellulose,
and complex substrates like maize, waste sulfite liquor from paper industry and
molasses [
Beesch, 1952
]. It is not surprising that consequently a huge effort was
spent on optimal media research.
A first production plant for acetone production was erected in 1915, in the later
years similar plants followed in Canada, India, France and the United States.
The rising of automobile industry around 1920 led to an increased demand of
butanol which was hitherto an unwanted by-product of acetone formation. In
parallel, petrol-industry developed, and competitiveness of the process became an
issue. This inspired research on cultures using starch more efficiently as energy
source. Little success was granted to this approach and research was diverted
to the investigation of fermentations on a multitude of monosaccharides from
hydrolysates of complex carbon sources. The found strain CSC no.8 was able to
ferment up to 6
.
5 % of the sugars, leading to a butanol yield of 2 %. Increasing
2.1. MOTIVATION OF RESEARCH 9
this yield became a major task in research. In particular for Great Britain that
was heavily depend on import, the different carbon sources became the limiting
factor. 60 % of the butanol production costs were caused exclusively by the
substrate.
Competitiveness of Bio-Butanol Lasted Until the End of World War II
The improvement of strains remained an issue until the 2nd World War where
the demand of acetone again drastically increased. As a result, semi-continuous
fermentations were operated with a multi-column continuous distillation down-
stream to extract alcohols. Until 1960 the use of fermentation as production
route virtually ceased in all Western Countries, a plant in South Africa remained
operational until 1983.
10 CHAPTER 2. INTRODUCTION
2.2 Biological Facts
This section will deal with the general introduction to the biology of C. aceto-
butylicum (2.2.1), which is then followed by a summary of biochemical production
pathways (2.2.2) and genetic optimisation of strains (2.2.3).
2.2.1 What is Clostridium acetobutylicum?
Clostridium acetobutylicum is a member of the firmicutes genus. It is an obligate-
anaerobe, Gram-positive and spore forming organism that is able to ferment a
variety of different sugars and convert them to acetic acid, butyric acid and solvents
as acetone, butanol and ethanol in the typical ABE-fermentation [Duerre, 2005].
Its genome sequence was recorded and annotated in 2001 [
Noelling et al., 2001
]. It
consists of one main chromosome (3
.
94
Mb
) and a mega plasmid pSOL1 (192
kb
),
which contain 3740 protein-coding open reading frames and 107 RNA genes.
The life cycle consists of three distinct phases [Luetke-Eversloh and Bahl, 2011]:
•
acidogenesis: In this phase the cells are exponentially growing and the
products acetic acid and butyric acid prevail.
•
solventogenesis: In this phase the cells take up the excreted acids and
metabolise them to the corresponding alcohols, ethanol and butanol, as
well as acetone, in a fixed ratio depending on the substrate. For glucose as
substrate the ratio of ABE products is 3:6:1.
•
sporulation: In this phase, productivity ceases and cells transform into a
durable state until environmental conditions ameliorate.
The major part of solventogenic genes is located on the pSOL1 plasmid
[
Grimmler et al., 2011
]. Losing the plasmid thereby results into a solvent negative
strain [
Rogers, 2002
]. In order to prevent this loss, it is required to apply several
stress parameters on the cultures e.g. addition of acids, decreases in pH, changes
in dilution rate or temperature [Barbeau et al., 1988].
2.2.2 Biochemical Pathways
Carbohydrate Uptake
There are two distinct uptake systems in bacteria, either ion channel mediated
uptake along an ion gradient, mainly H
+
and Na
+
ions, or active import by
cleavage of high-energy bonds, mainly ATP or phosphoenolpyruvate (PEP). The
latter of both mechanism is the predominant in clostridial species. A multitude
of sugars can be imported by different phosphotransferase systems (PTS), in
total 13 systems are known in C. acetobutylicum, including one on the pSOL1
2.2. BIOLOGICAL FACTS 11
plasmid [
Duerre, 2005
, p.155ff]. It was shown, that glucose uptake is pH dependent
[Yerushalmi et al., 1986b].
Glycolysis and Acid and Solvent Pathways
Biochemical studies on glycolysis are rare in C. acetobutylicum, the pathway is
mainly inferred from the genome sequence. Only the glyceraldehyde-3-phosphate
dehydrogenase has been analysed [
Duerre, 2005
, p.675]. However, the main path-
way for acid and solvent production are extensively studied. This section summar-
ises [
Duerre, 2005
, p.671ff]. As shown in figure 2.1, activation of pyruvate (Pyr)
Ac-P Ac-CoA
AcAc
AcON
Ac
Bu-CoABu-P
AcAc-CoA
Bu
Pyr
Bual
Acal EtOH
BuOH
La
Cro-CoA
BHBu-CoA
Glucose
CAC1742
(pta)
CAP0165 (adc)
CAC1743
(ack)
CAC3076
(ptb)
CAC2873 (thlA)
CAP0078 (thlB)
CAP0163
CAP0164
(ctfAB)
CAC3075
(buk)
CAC2229
CAC2499
(pfor)
CAP0162
CAP0035
(adhe)
CAP0162
CAP0035
(adhe)
CAP0025
CAC3375
CAP0035
CAP0162
CAC3298
CAC3299
(BdhAB)
CAC0267
CAC3552
(ldh)
CAC2711 (bcd)
CAC2708 (hbd)
CAC2712 (crt)
Glycolysis
Figure 2.1: Production pathways of butyric and acetic acid and the solvents
acetone, ethanol and butanol. Adopted from [
Duerre, 2005
, p.674] and
[Lee et al., 2008b]. Abbreviations are explained in the text.
to acetyl-CoA (Ac-CoA) is achieved via the pyruvate ferredoxin-oxidoreductase
(pfor). This system produces hydrogen and carbondioxide. Lactate (La) is pro-
duced via a lactate deyhdrogenase (ldh). Acetyl-CoA is the key compound in acid
and solvent production. Condensation of two molecules acetyl-CoA via thiolase
(thlAB) leads to one molecule acetoacetyl-CoA (AcAc-CoA). Acetoacetyl-CoA
12 CHAPTER 2. INTRODUCTION
is converted to
β
-hydroxy-butyryl-CoA (BHBu-CoA) that is subsequently de-
hydrated to crotonyl-CoA (Crt-CoA) which then is converted to butyryl-CoA
(Bu-CoA), via the three enzymes 3-hydroxybutyryl-CoA dehydrogenase, crotonase,
butyryl-CoA dehydrogenase (hbd,crt,bcd) respectively. These three genes form
an operon.
Depending whether acidogenic or solventogenic conditions prevail, the fluxes from
acetyl-CoA and butyryl-CoA are diverted into different directions: During acido-
genesis, each CoA-derivative is phosphorylated with inorganic phosphate by their
respective phosphotransferase, phosphotransacetylase or phosphotransbutyrylase
(pta and ptb), into acetyl-phosphate (Ac-P) and butyryl-phosphate (Bu-P). These
phosphates act as donor for the reaction of two distinct kinases, acetate kinase
(ack) or butyrate kinase (buk) to generate ATP and the acids, acetate and butyrate
(Ac and Bu).
During solventogenesis, acid re-uptake acts either by the reverse reaction of the
kinases, or via an acetoacetyl-CoA: acetate/butyrate-coenzyme A transferase
(ctfAB) consisting of two subunits. ctfAB accepts acetoacetyl-CoA as CoA-donor
and transfers the CoA to the acids, the products are the respectve CoA-acid de-
rivative and acetoacetate. Acetoacetate decarboxylase (adc) acts on acetoacetate
and produces acetone (AcON). The two CoA-derivatives are reduced to their
respective aldehydes (adhE) and alcohols via unspecific alcohol dehydrogenases
(BdhAB). A complete view on reaction mechanisms of these enzymes is given
elsewhere [Gheshlaghi, 2009].
Uptake of Acids
The investigation of the effect of propionic and acetid acid uptake on the
metabolic spectrum during batch fermentation showed that two pathways are
active, the CoA-transferase pathway and the kinase-phosphotransbutyrylase
pathway. The latter was dominant and led to an increase of solvent yields
[
Huesemann and Papoutsakis, 1990
]. Phosphotransbutyrylase acts on its sub-
strates in a ping-pong like mechanism. Its activity is highly sensitive to pH changes
in the physiological range, mainly in the butyryl-phosphate direction but not so
much in the butyryl-CoA direction. Further it is inhibited by ATP, suggesting a
role of ATP for the determination of reaction direction [
Wiesenborn et al., 1989
].
Both re-uptake paths of acids are confirmed by several pulse experiments in
batch culture. Flux balance analysis showed that in batch culture the short-term
response to an acetic acid pulse in acidogenesis, the flux from acetyl-CoA to
acetoacetyl-CoA is increased, but not the CoA-transferase activity. The long-
term response showed increased acetone and reduced butanol concentrations. A
butyrate pulse in chemostat culture showed that butyrate is mainly taken up
via the CoA-transferase pathway, however the butyrate synthesis still prevailed
[Junne, 2010].
2.2. BIOLOGICAL FACTS 13
The Metabolic Shift
The metabolic shift is the transitional phase C. acetobutylicum undergoes when
acidogenesis halts and solventogenesis starts. It is reversible [
Haus et al., 2011
]
and acidogenic conditions are maintained above pH 5.7 in continuous culture,
solventogenic conditions are established below a pH of 4.3 [
Janssen et al., 2010
].
The conversion of acids into solvents is seen as a de-acidification mechanism
[
Monot et al., 1984
,
Huang et al., 1985
]. Since the internal pH of C. acetobutyl-
icum cannot be maintained at a constant level [
Gottwald and Gottschalk, 1985
],
this suggests that the switch is linked to the pH. Indeed, in pH-uncontrolled cul-
ture a surplus in acid production or addition of high amounts of acetic acid (up to
200
mM
) cause an acid-crash which blocks solventogenesis [
Maddox et al., 2000
,
Cho et al., 2012
]. These results were confirmed by addition of less then 2
mM
formic acid in two separate studies [
Wang et al., 2011
,
Cho et al., 2012
]. The
importance of internal pH and the pH-gradient between cell and the reactor are
extensively discussed elsewhere [Papoutsakis et al., 1987].
As undissociated butyrate concentrations are linked to internal pH, its precurs-
ors like butyryl-phosphate may also have signalling function [
Desai et al., 1999
,
Paredes et al., 2005
]. Indeed a butyrate kinase deletion mutant showed a sig-
nificant earlier start of solventogenesis compared to the wildtype. Addition-
ally, the concentration of butyryl-phosphate was bimodal with one peak cor-
responding to solvent production and one corresponding to carboxylic acid re-
utilization [
Zhao et al., 2005
]. The onset of solventogenesis showed correlations
with butyryl-CoA spikes in batch culture [
Boynton et al., 1994
] and to undissoci-
ated butyric acid levels around 6
−
13
mM
[
Monot et al., 1984
,
Huang et al., 1985
,
Terracciano and Kashket, 1986
,
Huesemann and Papoutsakis, 1988
]. A current
study of a deletion mutant of the butyric acid production pathway re-evaluates
the hypothetical role of butyrylphosphate and butyryl-CoA and it comes
to the conclusion that neither is necessary for the onset of solventogenesis
[
Lehmann and Luetke-Eversloh, 2011
]. Similarly, the role of acetic acid on the
onset of the pH-shift is also controversial. Studies suggest both, that internal acetic
acid has no effect [
Bahl et al., 1982a
,
Zhao et al., 2005
], while acetate stimulus
experiments suggest it has an effect [Junne, 2010].
The alteration of electron flow is suspected to induce the shift [
Meyer et al., 1986
,
Rao and Mutharasan, 1987
]. Also a characteristic increase of the redox-potential
is observed during the shift [
Grupe and Gottschalk, 1992
,
Peguin et al., 1994
].
The deletion mutant of one of the two acid kinases and the CoA-transferase is
unable to produce solvents, the authors consider this a result of the inability to
control electron flow of C. acetobutylicum M5 [Sillers et al., 2008]
The DNA-topology is influenced by different environmental conditions and it
was shown that relaxation of the coiling increased acetoacetate decarboxylase
transcription [Duerre, 2005, p.680].
14 CHAPTER 2. INTRODUCTION
Finally, sigma factors are suspected to play a significant role, in particular Spo0A
phosphorylation has a primary role in the onset of solventogenesis. Spo0A deletion
mutants neither show sporulation nor butanol production [
Ravagnani et al., 2000
].
However, over-expression of this gene did not alter the onset time of solventogenesis
[
Alsaker et al., 2004
]. The general organization and regulation of solventogenic
genes in operons is reviewed by [Duerre et al., 2002].
In C. acetobutylicum P262, acidogenesis and solventogenesis seem to operate in
parallel - the cells undergo cyclic changes in productivity during a chemostat at
different dilution rates [
Clarke et al., 1988
]. The authors suggest, that it is un-
likely that both pathways are operating in parallel in one cell, a mixed population
hypothesis may be reasonable. Current research re-investigates this hypothesis
and models indicate such a possibility [Millat et al., 2013b].
2.2.3 Engineering a Butanol Production Strain
A summary of pathway remodelling approaches is given by [
Lee et al., 2008b
,
Luetke-Eversloh and Bahl, 2011
]. Butanol increase is reported by non-replicative
plasmid inactivation of butyrate kinase [
Green et al., 1996
]. Combination of the
butyrate-kinase mutant with over-expression of an alcohol dehydrogenase yielded
a strain that produced 16
.
7
g
L
of butanol [
Harris et al., 2000
]. A transcriptional
repressor for solvent synthesis was recognised and studied, its inactivation res-
ulted in a deregulated solvent production strain with better yields of solvents
[
Nair et al., 1999
]. Overexpression of the alcohol dehydrogenase and downregula-
tion of the CoA transferase using antisense RNA (asRNA) yielded an increase
of ethanol concentrations to 9
g
L
and butanol levels comparable to the wildtype
[Tummala et al., 2003a].
In a corresponding study, thiolase and alcohol dehydrogenase were overexpressed
by using the phosphotransbutyrylase promotor and again asRNA for ctfAB si-
lencing, a higher selectivity of butanol to acetone was achieved. Solvent titers
reached 30 g
L[Sillers et al., 2009].
Engineering of thiolase was performed using an E. coli library that was screened
for optimal thiolase activity before it was inserted into C. acetobutylicum. The
resulting strain showed less growth and optimised alcohol titers for both, ethanol
and butanol [Mann and Luetke-Eversloh, 2013].
Deletion and overexpression analysis of Spo0A revealed a fundamental role in
both, sporulation and initiation of solvent production as transcriptional reg-
ulator, which also regulates other sporulation factors, similar to B. subtilis
[Harris et al., 2002,Thormann et al., 2002].
The newly developed ClosTron insertion mutation II technology [
Heap et al., 2007
]
can be used to achieve highly reproducible knock-out mutants. Use of this sys-
tem is reported for transforming C. acetobutylicum into an ethanol producer
by inactivation of 3-hydroxybutyryl-CoA-dehydrogenase. Inactivation of this
2.2. BIOLOGICAL FACTS 15
whole pathway did neither alternate sporulation nor the onset of solventogen-
esis [
Lehmann and Luetke-Eversloh, 2011
]. Deletion of the phosphotransacetylase
alone does not change acetate production and the deletion of the acetoacetate
decarboxylase resulted in a drastically reduced acetone production. Combined
deletion of both genes increased the flux to butyryl-CoA leading to butyrate
[
Lehmann et al., 2012a
]. Inactivation of the acetate kinase alone did not alternate
the solvent production either, the double knock-out of butyrate and acetate kinases
is currently under investigation [
Kuit et al., 2012
]. The deletion of phosphotrans-
butyrylase showed as well high ethanol and butanol yields and a disruption
of butyrate production, pH control was necessary to allow the metabolic shift.
Metabolic profiles of mutants are pH-dependent: With no pH-control there is
no acetone, and accumulation of acetate. With pH-control there is acetone pro-
duction and re-uptake of acetate. It was equally shown that butyrate may be
re-assimilated in the ptb mutant [Lehmann et al., 2012b].
A different metabolic approach converts acetone to isopropanol: Insertion of
dehydrogenases from other species allow the production of more than 20
g
L
of
alcohol [
Lee et al., 2012
,
Dusseaux et al., 2013
]. Earlier, other authors reported
an increase from 2
g
Lto
18
.
8
g
L
by simultaneous disruption of both kinases and
the insertion of a mutated alcohol dehydrogenase [
Jang et al., 2012
], which is in
contrast to another study, who reported that their double knock-out strain did
not produce butanol [Sillers et al., 2008].
2.2.4 Concurrent Designs
Concurrent processes are established, the butanol production apparatus of
C. acetobutylicum is shuttled into E. coli and other bacteria. Although the
productivity is low, the authors suggest a high potential of this approach
[
Inui et al., 2008
,
Nielsen et al., 2009
]. Still, a complete unadapted organism
faces the same challenges of butanol stress on the cell membrane than the better
adapted Clostridia. A completely different approach to butanol production is
performed by employing monoxygenases and butane as substrate [
Duerre, 2005
,
p.685].On this process a patent is pending (patent EP 0 987 348 A1).
16 CHAPTER 2. INTRODUCTION
2.3 Butanol Fermentation
This section reviews the necessary technological and environmental parameters
that allow a sustained butanol production in C. acetobutylicum
2.3.1 Production of Butanol
Media Composition
A minimal medium with proteins and glucose supplemented with p-aminobenzoic
acid and biotin was sufficient to promote growth [
Beesch, 1952
]. When the fer-
mentation was operated in glucose or ammonia limited mode, no solvents but only
acids were produced. It was equally shown that butyric acid at a pH less than
5.0 could shift the culture to solventogenesis with improved ratios of products
[Bahl et al., 1982a].
The same authors showed that a phosphate limited medium could produce superior
results to the hitherto existent fermentations. They were able to ferment 54
g
L
of
glucose to 10 g
Lof butanol and 4 g
Lof acetone [Bahl et al., 1982b].
The role of ions in the fermentation broth was elucidated in the same year
by another group, who showed that iron, magnesium and potassium ions can
promote growth, only magnesium and manganese ions had a deleterious ef-
fect when applied in excess [
Monot et al., 1982
]. Conversely, it was shown that
iron limitation and viologene addition as redox-agent enhance the butanol yield
[Peguin and Soucaille, 1995].
Substrates, Product Yields and Product Spectra
The typical ratio of 6:3:1 of butanol, acetone, ethanol is reached on starch,
saccharose, xylose and fructose, while a ratio of 5:4:1 is reached on arabinose
[
Beesch, 1952
]. Using a mixture of glucose and xylose (1:1), it was found that
uptake of xylose seems repressed by glucose uptake, since xylose but not glucose
accumulated in the medium [
Fond et al., 1986
]. Solvent productivity on all the
three, glucose, xylose and its mixture was at most 0.8g
Lh , 0.58 g
Lh and 0.94 g
Lh .
Another mixed substrate fermentation of glucose and low-grade glycerol (1:1 and
2:1) in a chemostat showed that butanol was the major endproduct (43 % and 63 %)
and the culture could be maintained stable over 70 days. In two experiments, gluc-
ose was entirely consumed (15
g
L
and 30
g
L
) and 43 % of added glycerol was used.
Solvent productivity was assessed as 0
.
47
g
Lh
[
Andrade and Vasconcelos, 2003
].
Grown solely on glycerol, 1,3-propanediol is formed, when grown on rhamnose,
also 1,3-propanediol, propionic acid and propanol are produced [
Forsberg, 1987
].
Complex substrates like corn fibre yield lower solvent productivity (0
.
2
g
Lh
to
0
.
4
g
Lh
) and they require addition of xylanases to successfully start the batch
fermentation. Uptake of glucose and arabinose was superior to uptake of galactose
2.3. BUTANOL FERMENTATION 17
and xylose [
Qureshi et al., 2006
]. Similarly, a complex substrate mixture from hy-
drolysates of distillers grain was studied on different clostridial species. Clostridium
acetobutylicum could reach a solvent productivity in batch operation mode of
about 0.25 g
Lh , fermenting the following sugars in decreasing order of preference:
glucose, arabinose, galactose, cellobiose, xylose, mannose. Inhibitory effects of
the hydrolysis products syringaldehyde, ferulic and p-coumaric acid, as well as
the growth stimulating effect of furfural and hydroxymethyl furfural were demon-
strated in this study [Ezeji and Blaschek, 2008].
Input of carbondioxide to the fermentation by gassing inhibits dehydrogenase
activity and increases butanol yields [Kim et al., 1984]. In a glucose limited che-
mostat, carbon monoxide gassing leads to decreased growth rates but increased
glucose uptake, with no acetone but sustained butanol production of 0
.
74
g
Lh
[Meyer et al., 1986].
2.3.2 Counteracting the Effects of Butanol
Butanol Effects
The accumulation of butyrate and acetate in the membrane does not cre-
ate massive cell leakage [
Huang et al., 1985
]. Butanol however has a chao-
tropic effect on the membrane which results in early cell death and small pro-
ductivity. It is one major focus in optimisation proceedings. It was shown
that the internal pH could not be maintained and there was leakage of ATP
[
Bowles and Ellefson, 1985
,
Balodimos et al., 1988
] or PEP and therefore a stop
of the glycolytic flux [
Gheshlaghi, 2009
]. In contrast, a metabolome study showed
that addition of 5
g
L
butanol did not drastically affect intracellular metabolite
pools [
Amador-Noguez et al., 2011
]. This is confirmed by two butanol stress ex-
periments in continuous culture, a pulse experiment [
Janssen et al., 2012
] and
a stepwise forcing experiment [
Schwarz et al., 2012
]. Pulse experiments with
butanol in batch culture during the acidogenic and solventogenic phase increased
acetate uptake in the short term, but in the long term every reaction is reduced
[
Junne, 2010
]. As a result of butanol stress, the fraction of saturated to unsatur-
ated fatty acids showed a dose-dependent increase. Conversely, butanol challenges
of 0
.
25 % vol/vol and 0
.
75 % vol/vol were used to study the tolerance of two
strains, the pSOL1 deletion strain as control strain and the strain 824(pGROE1)
that contains the chaperone system groESL under a thiolase promotor. Butanol
addition increased the expression of the major stress responses and the solvent
formation genes, while it decreased the expression of genes for fatty acid synthesis
and glycolysis [Tomas et al., 2004].
A library enrichment study allows the identification of genes conveying butanol
tolerance. It was undertaken by transferring stationary phase cultures into media
with increased butanol concentrations. Strains containing a plasmid with the
18 CHAPTER 2. INTRODUCTION
CA
C1869
gene showed a 81% increase in butanol tolerance compared to the wild
type. It grew to higher cell densities and showed a prolonged metabolism. It
remained an open question which regulations were changed due to over-expression
of this gene [Borden and Papoutsakis, 2007].
Fermentation Operation Modes
A review on butanol toxicity and possibilities to overcome it, is given by
[Ezeji et al., 2010].
First attempts to reduce the chaotropic effects of butanol were started by using
liquid-liquid extraction with n-decanol saturated with butyric acid. A continuously
operated membrane bioreactor was therefore connected to a mixer-settler cascade.
A fourfold increase of butanol (0
.
51
g
Lh to
1
.
96
g
Lh
) could be noted. Direct contact
to the decanol phase caused cell damage however. Additionally, butyric acid
saturation of the extraction phase was necessary to prevent its removal from the
fermentation process [Eckert and Schuegerl, 1987].
Pervaporation, that is evaporation of a liquid after diffusion through a membrane,
was used to remove butanol in chemostat culture. It increased the product forma-
tion rate to 2.34 g
Lh and higher [Izak et al., 2008].
Perstraction is a similar process where butanol is allowed to diffuse through a
permeable membrane into an ionic liquid. An increase of butanol production from
0
.
057
g
Lh
to 0
.
21
g
Lh
was reached. An eightfold higher amount of lactose could be
fermented during this batch fermentation compared to the usual batch operation
mode [Qureshi and Maddox, 2005].
In all these methods the fermentation broth is in direct contact to the membrane
and fouling of the membrane becomes an issue. A method like vacuum product
recovery overcomes this problem. Although butanol has a higher boiling point,
the azeotropic mixture with the other alcohols in the fermentation broth leads to
a better vaporisation. This approach allows the cells to completely utilise glucose
for higher growth and higher concentrated product streams [
Mariano et al., 2011
].
Biocatalysis by immobilised cells allows the conversion of substrate to the desired
product, in a first attempt C. acetobutylicum was supplied with a medium that
did not support cell growth and effects of different feeding stream metabolites
were investigated. The packed-bed reactor was run in continuous mode, cells
were immobilised in alginate beads. Productivity of this system was lowered
by sporulation and cell death, still butyrate supply of 2
g
L
allowed produc-
tion of 1
.
9
g
L
butanol after 10
h
of cultivation. Activity regeneration in the
same system can be reached by supply of ammonia and vitamins in the feed
[
Reardon and Bailey, 1989
,
Reardon and Bailey, 1992
]. In a different setting, im-
mobilised cells were examined for their extracellular
α
-amylase activity on starch
in the feed flow. The total solvent yield was 1
g
L
[
Badr et al., 2001
]. Finally,
biomass recycling was used to increase butanol yields. It was shown that under
2.3. BUTANOL FERMENTATION 19
non-glucose limiting conditions recycling lowers the ATP demand and increases
solvent yields, under limiting conditions only higher yields of acids were reached.
In general, a range of total solvent productivity between 4
.
2
g
Lh
to 6
.
5
g
Lh
were
reached and the authors proposed an experiment that would ultimately lead to a
total solvent productivity of 12.4g
Lh [Meyer and Papoutsakis, 1989].
20 CHAPTER 2. INTRODUCTION
2.4 Published Data
This works aims at integration of transcriptome data and metabolome data and the
evaluation of database information in view of transcriptome data. A short introduc-
tion into both omics will be given here (2.4.1). For completeness, some proteome
studies are also mentioned. For a more in-depth introduction about opportunities
and pitfalls, refer to this review on bacterial omics [Mashego et al., 2007].
In the second part of this section, three different experiments will be introduced
in more detail (2.4.2). They will serve as data for the models that are going to be
established throughout this thesis.
2.4.1 Analysis Techniques for Different Omics
A short review on omics is given by [Fiehn, 2001,Joyce and Palsson, 2006].
Metabolomics
The physico-chemical properties of the ABE-products allow an easy detection
by gas chromatography [
Fond et al., 1984
,
Green et al., 1996
,
Harris et al., 2000
,
Lehmann et al., 2012a
]. However, since the evaporation of acids may impose
some methodological problems, the use of a High-Performance Liquid Chroma-
tography (HPLC) and a refractive index detector is also frequently encountered
[
Buday et al., 1990
,
Tomas et al., 2003
,
Tummala et al., 2003a
,
Kuit et al., 2012
].
The coupling of a tandem mass spectrometer to the HPLC allows the de-
termination of intracellular metabolites. This procedure requires several pre-
paratory steps, e.g. rapid sampling and rapid quenching [
Schaub et al., 2006
,
Schaedel and Franco-Lara, 2009
]. Such approaches were used for determining
intracellular metabolites of E. coli grown in
C13
-glucose supplemented medium
[
Schaub, 2005
,
Bajad et al., 2006
]. For C. acetobutylicum, one similar study of
a batch culture is published. 121 metabolites were measured using a tandem
mass spectrometry after addition of universally labelled
C13
-glucose. Massive
changes in all metabolites during the shift from acidogenesis to solventogenesis
were observed. The carbon flux is redirected from biomass growth to solvent
production [Amador-Noguez et al., 2011].
Online measurements procedures are published for metabolite analysis using a
mid-infrared spectroscopy approach [
Kansiz et al., 2001
] and for redox balance
determination using a fluorescent probe [Srivastava and Volesky, 1991].
Transcriptomics
The analysis of the complete transcriptome of C. acetobutylicum allows the tem-
poral resolution nowadays. Numerous such data sets are available: Study of Spo0A
overexpression [
Alsaker et al., 2004
], groESL overexpression [
Tomas et al., 2003
],
2.4. PUBLISHED DATA 21
ctfAB knockdown [
Tummala et al., 2003b
] and the transcriptional programme
of sporulation [
Alsaker and Papoutsakis, 2005
,
Jones et al., 2008
] were performed.
Responses to butanol addition [
Alsaker et al., 2004
,
Alsaker et al., 2010
,
Janssen et al., 2012
,
Schwarz et al., 2012
] and to several acids [
Alsaker et al., 2010
] were recorded. Re-
production of array results is usually undertaken by using a real-time PCR
approach on few genes [
Nolan et al., 2006
,
Lehmann and Luetke-Eversloh, 2011
].
The quantities measured by both approaches are in general comparable
[Dallas et al., 2005].
Proteomics
Stress response related proteins were detected using pulse-labelled proteins in
a batch culture [
Terracciano et al., 1988
]. The proteome study of a phosphate-
limited chemostat culture analysed 130 proteins and found 52 proteins being
up-regulated two-fold during the onset of solventogenesis, and 34 proteins being
downregulated by the same factor [
Schaffer et al., 2002
]. A more sensitive pro-
teome protocol was developed and tested in a similar culture, yielding a resolution
of over one thousand proteins on a 2D gel [
Schwarz et al., 2007a
]. In a phosphate-
limited chemostat, 15 proteins could be specifically assigned to acidogenesis and
29 to solventogenesis [Janssen et al., 2010].
2.4.2 Data Treated Within This Thesis
Three different data sets will be shown here, the standard batch fermentation in
complex medium, the acid stimulation experiments of that batch fermentation,
and the continuous fermentation under phosphate limited conditions.
The Standard Batch Fermentation
The first fermentation in which metabolomic and transcriptomic data were both
collected was done in the Papoutsakis laboratory [
Alsaker and Papoutsakis, 2005
,
Jones et al., 2008
]. The transcriptome data is large as it was collected over 25
samples along the whole fermentation time. A batch culture was grown in complex
growth medium (CGM) and maintained at pH
≥
5 until sporulation. Butyrate
spiked after 16
h
and butanol production started at the same time. The expo-
nential growth became stationary and cells switched to solventogenesis. This
fermentation yielded 11
g
L
of butanol after 45
h
. Cells continued to morphologic-
ally change until 60h but productivity ceased. A K-means cluster algorithm split
the transcriptional analysis data into five groups, corresponding to several phases,
thus identifying the genes significantly up-regulated during these phases.
During the first phase, cell motility genes are up-regulated, as well as glucose
transporter genes and nucleotide transporter genes.
The second phase is marked by an increase in energy production and the sporula-
22 CHAPTER 2. INTRODUCTION
tion relevant markers abrB and sinR are up-regulated.
The third cluster is overlapping with the second, however expression is sustained
over a longer period and genes relevant for fatty acid biosynthesis and iron import
are up-regulated. In this third cluster, the key solventogenic genes, Granulose form-
ation genes and stress and heat shock proteins are up-regulated. Also branched
amino acid synthesis seems up-regulated.
The fourth cluster contains numerous carbohydrate relevant uptake genes, and
genes encoding for transport of inorganic ions and amino acids. Starch metabol-
ism genes were also up-regulated and may be involved in granulose formation.
Additionally, arginine biosynthesis genes were up-regulated, because arginine was
probably depleted in the medium.
Acid Stimulus During Batch Fermentation
In the stimulation experiment of the Papoutsakis group, acetate, butyrate and
butanol were added during the exponential phase [
Alsaker et al., 2010
]. Preparat-
ory studies suggested that levels of 46
,
78
and
107
mM
of acetate, and levels of
17
,
33
and
49
mM
of butyrate negatively affect the metabolism. A summary of
all metabolic effects is given in the paper.
While acetic acid stress (45
mM
) down-regulates butyrate formation and vice
versa, butyrate stress (50
mM
) down-regulates acetate formation. Addition of
acids up-regulated stress response and solventogenic genes but unexpectedly not
sporulation relevant genes. Transporter proteins, post-translational modification
proteins and energy metabolism genes are up-regulated after acetate and butyrate
stress. The amino acid transport and metabolism show both, upregulation and
downregulation. Spo0A is slightly up-regulated, but upregulation ceased for spor-
ulation genes within 6
h
.Glucokinase and the IIABC phosphotransferase-system
were first highly expressed and later downregulated.
The Continuous Fermentation According to COSMIC-SOP
This section summarises the paper of. [Grimmler et al., 2011].
Acidogenic steady state conditions at pH 5.8 are reached after approximately
150
h
, 25
mM
of acetate are present and 51
mM
of butyrate. After switching
off pH-control, pH 4.5 is reached and then maintained actively by pH-control.
Solventogenic conditions are thereby established, and 37
mM
of butanol and
24
mM
of acetone are produced. Ethanol concentrations remained unchanged
in both conditions, 5
mM
and 8
mM
. Glucose showed a peak during the shift,
from 51
mM
under acidogenic conditions to 92
mM
during the shift to 86
mM
under solventogenic conditions. The peak of glucose corresponds to maximum
concentrations of acids and optical density.
245 genes were differentially expressed and collected in 4 groups, up-regulated
2.4. PUBLISHED DATA 23
genes under either acidogenesis (Group 1) and solventogenesis (Group 2), induced
(Group 3) or repressed (Group 4) only during the shift:
Group 1
arginine biosynthesis, flagellin biosynthesis, acetyl-CoA conversion to
crotonyl-CoA, alcohol dehydrogenase ,
Group 2
endoglucanases, glycerol-3-phosphate dehydrogenase, flavodoxin, cysteine
and sulfur metabolism, fatty acid synthesis, glycosyltransferases, solvento-
genic genes, cellusomal-like genes
Group 3
pyruvate decarboxylase, stress response, predicted mannose uptake
system
Group 4
carbon-monoxide dehydrogenase, glycerol-3-phosphate transport, tri-
carboxyacid cycle.
The increase of glucose during the shift is argued to reflect the increased carbon
uptake of the organism through acid re-assimilation. Sporulation and solvento-
genesis are two separate events and also the transcription of stress related genes
was already initiated when butanol was not present in the medium. Also, fatty
acid synthesis may not be a consequence of butanol stress, as the relevant genes
are up-regulated already during the transition.
24 CHAPTER 2. INTRODUCTION
2.5 Summary and Thesis Proposal
The necessity of researching alternative production routes for fuels in particular
for butanol is widely accepted. An effective and competitive process would not
only sustain the transition from fossil fuels to regenerative energy but it would
equally offer independence of a national economy to imports of raw oil. Since
from the historical perspective C. acetobutylicum was primarily an acetone produ-
cer, butanol production was not in focus. By considering the physico-chemical
properties of butanol in combination to the lengthy optimisation approaches to
overcome problems of process design, the design of an optimal production strain
is a very demanding task and a serious disadvantage to this process. The different
possibilities of genetic engineering propose a new approach to overcome the before
mentioned problems.
In silico design of an organism helps in reducing the experimental workload
associated with this engineering problem. It can suggest experiments on the one
hand, and it procures the scientist with a tool to investigate and focus on key
aspects of the system on the other hand. This thesis contributes to such a design
by proposing approaches to the following tasks:
1.
Investigation of data with a focus on the relationship of the reactome with
the underlying proteomic or transcriptomic data profiles.
2.
Hypothesis procurement for the filling of annotation gaps in the proteomic
database of C. acetobutlicum based on database and data information.
3.
Generation of a model that integrates transcriptome and metabolome data
for the in silico description of different culture and process designs.
4.
Hypothesis procurement for the directed generation of mutants from the
estimated parameters of the dynamic model.
2.6. OUTLINE 25
2.6 Outline
Chapter 3provides answers for the first two tasks. It will present an automated
modelling approach that integrates reaction annotations and temporal transcript
level data from several experiments into a database. The evaluation of transcript
data from a reactome point of view adds a further dimension for data evaluation
besides gene annotation and may enhance readability of the data. From a graph-
theoretic formalism approaches to investigate the transcriptome data are proposed.
Reduction of complexity and several examples will be shown. Missing reaction
annotations represent gaps in this database. The investigation how to find them
and infer knowledge will occupy the second part. A comparative approach will
be proposed that integrates again transcriptome data and database-information.
This approach will be carried out for a specific enzyme, the 3-hydroxybutyrate
dehydrogenase which is not annotated in C. acetobutylicum but in numerous other
organisms.
Chapter 4provides answers to the other two tasks by integrating time-series
of transcript level data into a reaction network model of clostridial butanol syn-
thesis. The model construction and model evaluation with respect to metabolomic
data and mutation experiments occupies this chapter’s first part. It will be shown
that the metabolic shift can be solely explained through transcript dynamics
without requiring considerations of pH. The second part then treats local and
global sensitivity analysis and their use for finding an experimentally feasible
optimal parameter sets that increases solvent productivity, which will be presented
for several reactions.
Chapter 5focusses on one implementation problem of this novel dynamic model
type. High numerical effort for the calculation of the integrated transcript level
profiles needs alleviation. Compression using the principal component analysis
does not only allow to increase calculation speed, but it will also inspires a novel
tool for model analysis in scope of optimisations by varying dynamic pattern
in the data, and it will inspire a data analysis routine in scope of clustering of
regulatory similar information.
Chapter 3
Automated Network Model
Creation
The important thing in science
is not to so much to obtain new facts
as to discover new ways of thinking about them.
Sir William Bragg
Biological information tends to be very heterogeneous in its qualities: Tran-
scriptome data stand beside the network structure of a biological pathway, protein
sequences stand beside regulatory cascades. However, these information represent
one same organism and belong together. Their integration represents one key
challenge. This chapter proposes an automated procedure to answer some aspects
of this challenge.
Starting from an overview of existing approaches of available pathway-databases
for C. acetobutylicum (3.1), it is deduced that an independent and more flex-
ible solution is required here. This database then is extended by integrating
experimental transcript level time-series in a novel way (3.2). Some examples
of derived models will be given in a graph format (3.3). Gaps in the present
annotation represent a challenge that this model cannot cover. A strategy to
overcome such gaps is annotation-transfer between species for which a methodo-
logy will be proposed (3.4). This will lead to the construction of hypotheses that
can be experimentally verified and a second possibility to integrate a different
type of data-base information, the enzyme annotation and Pfam-motifs, with
transcriptome experiments. As case study the research of a 3-Hydroxybutyrate
dehydrogenase in C. acetobutylicum will be undertaken (3.5). The conclusion is
given in section 3.6.
28 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
3.1 Database-Harvesting
3.1.1 Existing Pathway-Models
A metabolic network, synonymously a pathway-model, is a mapping of compounds
to compounds via reactions. Reactions are the result of an enzyme activity, which
is the result of the protein product of a specific gene. The correct gene-reaction
annotation is an ongoing process in the databases, e.g. not all reactions bear E.C.
numbers, not all genes are identified as enzymes. Many tools are available to
bridge some of such short-comes [Durot et al., 2009].
The online database MetRxn was only recently published. It harvests information
from KEGG, BRENDA and MetaCyc. Extensive curation effort was put in this
database by resolving naming inconsistencies, by balancing of stoichiometry and
charges, and by reconciliation of missing information between the databases. The
aim of MetRxn is the creation of pathway-models usable for flux balance analysis,
the comparison of conserved reactions throughout species and the identification
of all possible conversion routes from one substrate to one product. In total, 44
models of organisms are stored there [
Kumar et al., 2012
]. Notwithstanding the
efforts, download of the pathway-model of C. acetobutylicum from MetRxn was
not possible at writing time of this thesis and it could not be used. It consists of
430 metabolites connected by 363 reactions .
Earlier, a similar network was created and manually curated, containing 479
metabolites connected by 502 reactions [
Lee et al., 2008a
]. It aimed again at
flux balance analysis and growth performance evaluation from fermentation data.
Curation of the network was performed by several gap-filling procedures, as e.g.
BLAST research of missing enzymes. Single gene deletion studies suggest that
194 reactions are essential, and 27 reactions are partially essential. Parallel to this
work, other authors published their network [
Senger and Papoutsakis, 2008
]. It
contains 422 metabolites in 552 reaction. Curation was performed using a maximal
flux criterion for biomass production in conjunction with biomass constituting
equations and a thermodynamical consideration of the free Gibbs energy. Flux
balance analysis was again the aim and outcomes of single gene knock-outs were
proposed.
All three authors made neither workflows for database creation nor the databases
themselves publicly available. A reactome-knowledgebase has been established for
various other organisms and allows integration of softwares, namely Cytoscape
and R, as well as the interpretation of transcript expression data on the website
[
Matthews et al., 2009
]. In the commercial software Insilico a published model
for C. acetobutylicum exists that contains 507 reactions and 310 compounds.
3.1. DATABASE-HARVESTING 29
3.1.2 Local Solutions
Published networks were not available and their construction tools were not
accessible. Also the designation published softwares was very specific and did not
suffice for several features that will be required later.
The first task for any local solution is to gain access to a database, e.g. KEGG.
While there are several published softwares that access KEGG directly (e.g.
YANAsquare [
Schwarz et al., 2007b
]), again none of them seems flexible enough to
deal with the manifold of deposited information, nor was it possible to download en-
tire genomes. A programme used for building pathway-models and integrate data is
the very recently published RAVEN-toolbox [
Agren et al., 2013
]. Also the Vanted
toolbox integrates data into SBGN models from KEGG [
Junker et al., 2006
]. How-
ever, the most flexible solution for this thesis was Taverna: A programme able to
integrate a manifold of web-services, connect data-pipelines to one desired output
[
Hull et al., 2006
]. In combination with MATLAB it is possible to pipeline the
information for further use. Very recently, the creation of precise signalling models
from KEGG has been treated and a KEGG translator for networks published
[Wrzodek et al., 2013].
30 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
3.2 Network Analysis of the Clostridial Reactome
It is the scope of this section to create an on-place database that allows easy
information retrieval and its manipulation, and human-readable maps of the
stored annotations.
Through Taverna it was possible to automatically create a mapping from gene
identifiers (cac:) to reaction identifiers (rn:) to reaction partners (cpd:) by
harvesting KEGG with a workflow (B.1).
For the integration of transcriptome data to the downloaded reactions in a graph-
based format, a graph-theoretical backbone is introduced: Necessary notations and
the definitions of a graph are given (3.2.1) and a suitable software searched (3.2.2).
Structural characterisation of a biochemical network will aid in visualisation as
well as model analysis 3.2.3. Finally, a novel mathematical formalism will be
presented to integrate data into the database (3.2.4).
3.2.1 Graph Definition and Notations
Several possibilities to describe a metabolic network exist. One could define
reactions and metabolites as nodes and link both when the metabolite is part
of the reaction. Using a line graph transformation, one could as well study the
reactions linked by their respective metabolites [
Nacher et al., 2005
]. Here graphs
with only one type of nodes are treated: metabolite nodes that are being connected
with each other if and only if they are substrate and product of each other in
the same reaction. This graph will be named the metabolite-metabolite mapping
(MMM) or in mathematical notation, it is a graph
G
= (
V,E
) that contains two
sets, a set
V
of nodes
v
connected by a set of edges
e∈ E
. For convenience, the
genome of C. acetobutylicum is understood as set
X
of increasing numbers until
the last gene with number NJ:
X:= {j;j= 1...NJ}(3.1)
The mapping of the numbers in
X
to the corresponding reactions will be called
Rct
.
Rct
is not injective, since one reaction may be performed by several enzymes.
The graph
G
is generated by applying
Rct
on
X
. It is implicitly understood that
no node without edge must exist and that the set of nodes is given at all times
(e.g. all annotated metabolites in KEGG). Rct only acts on the edges:
GX= (VX,EX),EX:= Rct(X).(3.2)
Further, ”
∩
” and ”
∪
” denote the intersection and the union, respectively, and ”
\
”
is the set difference.
A transcript level at time tiof gene j∈ X will be named xj(ti).
3.2. NETWORK ANALYSIS OF THE CLOSTRIDIAL REACTOME 31
3.2.2 Softwares For Visualisation of Graphs
Visualisation of large networks and their annotations is a persistent problem
throughout systems biology. KEGG pathways are manually drawn and curated
and thereby easy to overview, however the greater scope of one metabolite within
the whole networks is lost. In order to gain a broader view on metabolite
connectivity, the first ansatz was to find a software that makes annotations and
pathways human-readable. Requisites to the softwares were: easy inter-operability
with MATLAB and Excel, ease of layouting, and visualisation of multidimensional
annotations. These softwares were tested:
•BioLayout Express 3D (BL3D)
•yED
•Cytoscape
•CellNetAnalyzer (CNA)
BL3D [
Theocharidis et al., 2009
] serves mainly for graphical layouting in two or
three dimensional space, it further offers two handy applications. One is the
possibility to search for a node in the internet by simple clicking it and accessing
a user-defined web page. The dbget functionality of KEGG can be perfectly
used for such purpose for all three types of identifiers (cac:,rn:,cpd:). Rapid
annotation retrieval in KEGG, makes BL3D an excellent tool for storing and
investigating graphs. A second important tool is the MCL-clustering algorithm
that detects densely connected nodes [
van Dongen, 2000
]. This algorithm will be
used in chapter 5. No interaction with Excel is possible.
yED has superior abilities compared to BL3D in aligning nodes and edges via a
huge library of algorithms. Graphical manipulation of node properties allows a
virtual designing of graphs in two dimensional space. Drawing of new nodes and
connecting them to the network is made easy. Uploading of annotations is not
possible. Again no interaction with Excel is possible.
Cytoscape’s [
Shannon et al., 2003
] essential strength is graph theoretical analysis
of network properties and still a large library of algorithms for graph formatting.
The possibility to integrate various annotations makes this the most power-
ful working tool [
Troyanskaya, 2005
,
Joyce and Palsson, 2006
]. Interaction with
Excel-files is implemented.
Finally, CellNetAnalyzer [
Klamt et al., 2007
] is a tool for the mathematical in-
vestigation of pathway models, its capabilities for visualisation are limited. Since
it is programmed in MATLAB, interaction with other scripts is facilitated.
Interaction of these four programmes with each other is complicated, while the
graphml-languages is supported by the first three softwares, CellNetAnalyzer ac-
cepts SBML, which the others do not. Still, the current version of graphml seems
32 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
not be standardized at such point that several softwares can be automatically
interlinked. This problem is reported for several other tools that use SBML as
well [
Joyce and Palsson, 2006
]. For this reason, interaction is granted by using a
simple node list. This however, makes re-layouting necessary. For this reason, the
use of Cytoscape is preferred because it offers a vast library of algorithms.
3.2.3 Graph Characterisation
A human-readable graph is achieved by visualising graph parameters, e.g. dis-
tance measures. This step also helps in characterising the network further
[
Stelling et al., 2002
,
Klipp et al., 2004
]. This approach focusses on certain
network properties, e.g. centrality measures give a prioritisation of targets
[
Aittokallio and Schwikowski, 2006
]. The following enumeration of network para-
meters relies on [Barab´asi and Oltvai, 2004].
Node Degree
The node degree is the number of neighbours of
v∈ V
. In a random network
the node degree probability
P
(
k
) of a node having exactly
k
connections is a
gaussian function. In a biological network this probability has a sharp peak at the
beginning and then it falls according to a power law:
P
(
k
)
∝kγ
with
γ <
0. Such
a network is called scale-free. The parameter
γ
further characterises the networks
robustness: For
−
3
< γ < −
2 the emerging network properties are robust against
failure of single nodes. For
γ≈ −
2, highly connected nodes are in contact with
the major parts of all nodes, while for
γ≈ −
3 these highly connected nodes
disappear and a random network emerges. Here, it is obtained by a non-linear
regression routine of Cytoscape.
Network Diameter
The network diameter ∆
G
is the maximal distance between any two nodes
v, w ∈ V
.
It serves as a measure for the compactness of the graph. However, only a small
diameter is a reliable parameter since it truly shows that nodes are within close
proximity, whereas a large diameter only shows that two nodes are distant,
while the others may be compact. Compact networks suggest an easy and rapid
communication of the interlinked nodes.
Centrality Measures
The eccentricity
Cecc
(
v
) is the reciprocal the shortest paths with maximal lengths
from the node v∈ V to other nodes w∈ V.
Cecc(v) := 1
max{dist(v, w) : w∈ V} (3.3)
3.2. NETWORK ANALYSIS OF THE CLOSTRIDIAL REACTOME 33
Thus, a high eccentricity shows that all other nodes are in proximity, whereas a
low eccentricity means that at least one node and its neighbours are very far.
The closeness
Cclo
is the reciprocal of the sum of all shortest paths that contain
v∈ V.
Cclo(v) := 1
Pw∈V dist(v, w)(3.4)
Likewise the eccentricity, high values are positive in the sense of proximity. The
closeness gives a tendency how the node is embedded in the graph, if either
isolated or central.
The radiality
Crad
is the average of the difference of the graph diameter and the
shortest paths from v∈ V to all other nodes.
Crad(v) := Pw∈V∆G+ 1 −dist(v, w)
n−1(3.5)
Hence, by consequently subtracting the shortest paths, the radiality becomes high
if all the paths are short, the node is then in the centre. Conversely, if all the
paths are long, then the node is in the periphery.
The stress centrality
Cstr
stands for the number of shortest paths
σst
from any
nodes s∈ V, t ∈ V different to v∈ V passing through v.
Cstr(v) := X
s6=v∈V X
t6=v∈V
σst(v) (3.6)
In biological terms, high stress shows how much a molecule is involved in the
cellular processes, it may not however symbolise how much this node is necessary
to hold together the different parts of the graph.
Likewise the stress centrality, the betweenness
Cbet
considers shortest paths from
nodes
s
to
t
passing through a node
v
, however it weights this number with the
total number of shortest paths connecting
s
and
t
, but not necessarily passing
through v.
Cbet(v) := X
s6=v∈V X
t6=v∈V
σst(v)
σst
(3.7)
Thereby, if
v∈ V
is the only connection between
s∈ V
and
t∈ V
, it gets a high
betweenness value. As complementary information to the stress centrality, this
value allows to assess the importance of a node to connect different parts of the
network.
A visualisation of two such parameters is easy to fulfil in Cytoscape: The node
colour and the node size are mapped to desired continuous graph parameters.
Further mappings are possible, e.g.discrete graph parameters can be mapped to
the node shape.
3.2.4 Data-Driven Network Generation - Methodology
The database that integrates transcriptome data and pathway information will
be called a data-driven pathway. By this integration, analysis of data is possible
34 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
by a perspective from the reaction network. Reduction of the whole network of
achievable reactions to the achieved reactions will help in characterising the current
status of the cell. It was suggested that the visualisation of networks during
different metabolic states provides a beneficial analysis tool [Khatri et al., 2012].
Surprisingly, the integration of short time-series data into networks for this purpose
is not encountered in literature, consider the review by [Dutta et al., 2009].
Boolean Rules for a Two-State System
The starting point is the entire graph as derived from the KEGG database
GKEGG
CAC = (V,E),E=Rct(X).(3.8)
GKEGG
CAC consists of 792 reactions that are connecting 852 metabolites.
For data-integration, a filtering approach will be used, so that pathway activity can
be assessed from the sub-graphs of
GKEGG
CAC
[
Aittokallio and Schwikowski, 2006
,
Reed et al., 2006
]: Knowing transcriptome experiments from several culture states
one can evaluate the bacterium’s regulatory events and provide hints on the cur-
rent necessity of enzyme synthesis and hereby activation of the conversion from
substrates to products. More precisely, one analyses two non-overlapping culture
states, e.g. acidogenesis vs solventogenesis or short term response vs long term
response. For simplicity they are referred to as s1and s2.
Xbu
(
s
) will denote the set of all transcript level expression values larger than some
boundary buat state s:
Xbu(s) := nj:xjts> bu, j ∈ Xo(3.9)
and
Xbl
(
s
) is the set of all transcript expression values smaller than some boundary
blat state s:
Xbl(s) := nj:xjts< bl, j ∈ Xo.(3.10)
Having this partition, a third partition is immanent, the set transcripts of which
is neither clearly repressed nor clearly induced, they are uncertain:
Xbl
bu(s) := j:bl< xj(ts)< bu, j ∈ X.(3.11)
Creation of Data-Driven Pathways From Logical Rules
Application of the three rules creates a boolean network from the initial graph
GKEGG
CAC . Two rules are shown here for state s2:
1. genes up-regulated at s2:Xbu(s2)
2. genes down-regulated at s1and uncertain at s2:Xbl(s1)∩ X bl
bu(s2)
3.2. NETWORK ANALYSIS OF THE CLOSTRIDIAL REACTOME 35
For simplicity, symmetry of the boundary parameters is assumed,
kblk
=
kbuk
=
b
.
While the first logical rule is the intuitive approach when considering transcript
data - up-regulation is considered as activation, the second rule transfers the
available information from one state to the other. If indeed expression were first
repressed during one state and it were relieved during the later state, it appears
that if a repression were relieved. Vice versa, if it were repressed in the second
state and uncertain during the first state, the organism appears to start repression.
This is an augmentation of state
s2
with respect to
s1
, because data from the
uncertain regions is used that would have been otherwise neglected. By this
augmentation is a positive statement about cell efficiency.
In order to distinguish the outcomes of these two rules, two types of graphs are
generated, the augmentation graph (
Ga
-graph) and the induction graph (
G
-graph).
For simplicity of evaluation a combination of both graphs, the
H
-graph, is suitable
H:= Ga∪G= (V,Ea∪ E).(3.12)
Note, that one must not augment when as reference an untreated culture is
used, as e.g. in [
Alsaker et al., 2010
], since augmentation only makes sense when
the reference state for microarray hybridisation is taken from the same culture,
either at a separate time-point or as average over all time-points. An external
reference is uninformative.
The following example illustrates the use of these rules to distinguish between
two states, e.g. acidogenesis (
s1
) and solventogenesis (
s2
). The non-augmented
graph Gduring s2reads:
G(s2) := V,E ∩ E(s2),E(s2) = RctXb(s2).(3.13)
Consequently, the augmented graph Hduring state s2then reads as
H(s2) := V,E ∩ E(s2),E(s2) = RctXb(s2)∪X−b(s1)∩ X−b
b(s2).(3.14)
Similarly, the active reactions during s1after augmentation are defined as:
H(s1) := V,E ∩ E(s1),E(s1) = RctXb(s1)∪X−b(s2)∩ X−b
b(s1).(3.15)
Validation and Model Reduction
Considering the two sets of edges
E
(
s1
) and
E
(
s2
) closer, it becomes apparent that
redundancy in biochemical pathways makes the intersection of both a non-empty
set despite the underlying sets of genes being distinct. It will therefore be of
interest to consider a third graph:
GcX:= V,E ∩ E(s1)∩ E(s2).(3.16)
36 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
As two entirely different states are compared, this graph will help identifying
sustained reactions. Finally, the graph of all inactive reactions will be necessary
for validation:
Ginact := V,E \ E(s1)∪ E(s2)(3.17)
Further reduction of graph size may become necessary to increase readability. One
way of doing this without loosing connectivity information, is to assume again
efficiency. A biological reaction from a substrate to a distant product takes a
shortest path without deviating to far distant molecules. Betweenness or stress
values larger than zero indicate that a node is used by at least one shortest
path. Thereby, iterative elimination of nodes from
G
with no stress value reduces
the network size. Ultimately, this means that solitary linear branches are cut
one-by-one from the periphery of the network until a bifurcation is reached.
3.2.5 Summary
This section proposes two rules for the creation of a pathway-model that integrates
transcriptome data with database knowledge to increase readability of data.
Use of the Data-driven Pathway
Literature shows that creation of meaningful flux balance network models is a
laborious task because of the urgent need to create a realistic representation of
the in vivo fluxes [
Lee et al., 2008a
]. There is no need to re-investigate such a
model, because several flux balance models for C. acetobutylicum do already exist
[Lee et al., 2008a,Senger and Papoutsakis, 2008].
The established data-driven pathway formalism serves for data visualisation in a
graph-based format. It was shown that regulatory networks are better analysed
in such a format [
Freeman et al., 2007
]. Therefor it was not urgently necessary to
unify compound isomers and to determine reaction directions, or to fill gaps that
are due to database errors. This explains the huge differences in metabolite and
reaction number between the published models and the raw model downloaded
from KEGG (900 in this study compared to 400 to 500 in other studies).
Boolean Rules for Integration of Omics
Some authors do report the application of boolean rules, but these rules are
not given [
Duarte et al., 2007
]. Other authors specify their rules in a one-
point measurement, however they do not include how they deal with con-
tinuous data [
Covert et al., 2001
]. Metabolic fluxes after deletion of unex-
pressed transcripts were used to relate transcript profiles and reaction profiles
[
Akesson et al., 2004
]. Pathway models were generated from up-regulation assess-
ments [
Patil and Nielsen, 2005
]. They proposed to track differentially expressed
genes but made no use of repressed genes or the temporal structure of their data.
3.2. NETWORK ANALYSIS OF THE CLOSTRIDIAL REACTOME 37
Projects like Reactome offer different clustering and data analysis algorithms, but
not the type proposed here.
Augmentation as proposed here unveils information from a two-point comparison.
Repression during one state and uncertain levels during a second state represent
additional information to the usual events of up-regulation. This additional
information is expected to reveal further information, as e.g. on constitutively
expressed genes.
Boolean Rules Can be Expanded to Multiple States
Although a comparison of two different states already offers valuable information
[
Reed et al., 2006
], this is a limitation that can be overcome easily. If there are
more than two clearly distinguishable states or phenotypes, then splitting the
time-series data into regions and using the correspondence of genes to these regions
helps in robustification of the model. For standardisation of transcriptome data, a
similar approach was suggested [
Yang et al., 2003
]. That is, a gene is counted as
up-regulated or down-regulated if this occurs more than once in the corresponding
region.
Rules and Data are Not Limiting
Boolean approaches has been shown to construct networks that contain a rich
complexity to be studied [
Dhaeseleer et al., 2000
]. The here presented approach
takes transcriptome data. In a similar fashion it could also make use of proteome
data. Further, the use of transcriptome data is not limiting. C. acetobutylicum is
not a fast growing organism, and it was shown that a mapping between transcripts
and proteins is possible in a single cell study if samples are not taken directly
after cell splitting [Golding et al., 2005].
Promising Results by Integrating Omics Data
Different Omics are already reported to be integrated into pathway models:
Integration of metabolome measurements from stimulus-response experiments
into pathways was treated by [
Cakir et al., 2006
]. They achieved the unification
of metabolome and transcriptome-measurements which enabled them to assess
whether genes are hierarchically or metabolically regulated. A successful approach
of transcriptome integration is reported: A graph was generated from genes that
were connected based upon Pearson correlation. Unfortunately, the biological
meaning of strongly correlated transcript expression profiles to gene-gene corres-
pondence is not discussed. The built model aided in mapping MCL-generated
clusters to specific tissues [Freeman et al., 2007].
In contrast to hypothesis-driven research, work with huge data from transcriptome
experiments allows knowledge discovery [
Bassett Jr et al., 1999
]. In this sense the
38 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
result of the integration of time-series data of transcript levels into a pathway
model is a tool that should not be underestimated. Here, the visualisation of
transcriptome data will eventually lead to knowledge discovery, as will be shown
in the following sections.
Outlook: Further Integration Possibilities
This pathway can be further filtered by integrating more information: The en-
zyme’s substrate specificity adds one filter criterion - substrates with a small
specificity can be neglected, and the number of links between two metabolites
effectively reduced. BRENDA offers such information and is readily accessible
in Taverna through a SOAP-service [
Chang et al., 2009
]. This route was not
undertaken here: Many enzymes have not yet been tested experimentally in C.
acetobutylicum, hence also these information require comparison approaches, e.g.
filling gaps by considering phylogenetically close relatives.
Reactions thermodynamics further discriminate the reaction directions. Fur-
ther, implementation of this approach would include redox-potential consid-
erations, intracellular pH measurements and energy balance determination
[Kumar et al., 2012]. Yet, available data are insufficient for such purpose.
Outlook: Possibilities of Validation
Viability of the network is one type of network validation commonly suggested
[
Reed et al., 2006
]. Building a data model from transcriptome data alone natur-
ally is not sufficient to create a viable organism, as only regulatory events are
detectable and constitutive genes are missing [
Troyanskaya, 2005
]. Validation
can still be carried out on the experimental level, using knock-out mutants and
phenotype comparison, enabling a check whether database entries are missing
[Reed et al., 2006].
3.3. DATA-DRIVEN PATHWAYS 39
3.3 Data-Driven Pathways
With the available integration scheme it is now possible to re-consider published
data. First, a possibility to fix a suitable boundary parameter
b
must be found
(3.3.1), then three experiments are introduced: the standard batch fermentation
in complex medium (3.3.3,
Gbatch
, [
Jones et al., 2008
]), the acetic acid addition
experiment in batch culture (3.3.4,
Gbatchpulse
, [
Alsaker et al., 2010
]), the pH-
shift experiment in continuous culture under phosphate limitation according to
COSMIC-specifications (3.3.5,Gconti, [Grimmler et al., 2011]).
For visualisation purposes the node for H
2
O with all its connections was deleted.
An overview of the considered states is given in table 3.1. Further, a summary of
some graph properties is given in table 3.2. The
γ
-parameter shows that the initial
Table 3.1: Considered states of data-driven networks in three published experi-
mental settings.
Gbatch Gbatchpulse Gconti
s110h 15min post pulse pH 5.8 (acidogenesis)
s240h 20h post pulse pH 4.5 (solventogenesis)
network and its data-driven derivatives are organised in communities. Although
these networks are not scale-free by common definition, the different filtering
approaches, both rules, and different boundary parameters do not drastically
change this parameter. Only for the inactive reaction graph it can be noticed
that no organisational structure is preserved, which is expected.
3.3.1 Derivation of the Boundary Parameter
For the determination of the boundary parameter, this section proposes to access
a structural property, the fraction of edges to nodes (E/N). This number allows
to track whether constitutive edges or peripheral edges are being eliminated when
a stricter, increasing, boundary parameter is used. Deletion in the periphery will
more likely create solitary nodes that are not considered further. In the converse,
a decreasing boundary parameter allows to assess whether new nodes are added
to the graph or existing nodes interlinked. This entirely refers to the boundary
parameter determination shown in figure 3.2.
Different Metabolic States are Distinguished by the Edges to Nodes Fraction
To deduce which states should be considered as
s1
a scan over ranges of
b
is
necessary. An example of such a scan over is shown for the stimulated batch
experiment in figure 3.1: Stress induced through the acetate pulse becomes
apparent here, only few genes are up-regulated directly after the pulse (
t1
). The
long-term response shows a steady increases of metabolic active genes until this
40 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
Table 3.2: Summary of generated graphs
none: no regression of γwas possible because R2≤0.6.
Graph bNodes Edges γ(R2) ∆G
GKEGG
CAC -902 2176 -1.29 (0.81) 12
Gconti(s1) 1.1 280 384 -1.33 (0.79) 15
Hconti(s1) 1.1 346 475 -1.35 (0.75) 16
Gconti(s2) 1.1 362 523 -1.51 (0.86) 14
Hconti(s2) 1.1 502 910 -1.40 (0.83) 14
Gconti,cX 1.1 144 181 none 11
Hconti,cX 1.1 165 204 none 13
Gconti,inact 1.1 400 355 -2.24 (0.97) 19
Hconti,inact 1.1 231 164 -2.65 (0.96) 7
Hbatch(s1) 0.8 306 447 -1.48 (0.78) 17
Hbatch(s1) 1.3 211 292 -1.42 (0.71) 18
Hbatch(s2) 0.8 394 585 -1.38 (0.77) 16
Hbatch(s2) 1.3 197 263 -1.41 (0.76) 16
Hbatch,cX 0.8 68 82 none 4
Hbatch,cX 1.3 46 61 none 4
Hbatch,inact 0.8 320 257 -2.57 (0.97) 11
Hbatch,inact 1.3 574 504 -2.70 (0.92) 22
Gbatchpulse(s1) 1.1 70 76 -1.59 (0.79) 6
Gbatchpulse(s2) 1.1 354 519 -1.40 (0.82) 15
Gbatchpulse,inact 1.1 512 465 -2.50 (0.92) 15
number attains its maximum at
t8
. Accordingly, reaction and metabolite numbers
follow this course. Major differences only become visible when the edges to nodes
fraction is calculated. The network which is most invariant to changes of
b
is
found at
t4
. Choosing
t1
and
t7
as representative time-points for two different
metablic states, the difference between both networks is maximal, and they are
clearly distinguishable from the network at t4.
Choice of b- Discrimination by Defects of E/N
The fraction of edges is expected to monotonically decrease, otherwise a defect
has occurred: The centrality measures of the initial graph
GKEGG
CAC
(figure 3.3)
indicate that the major part of the nodes is densely connected - betweenness
values are small - and there is not a large set of nodes lying in the graphs periphery
- closeness values are small. Additionally, by application of the boolean rules
the graph diameter decreases for each experiment but
γ
stays unaltered. These
observations show that the graphs integrity is loosened by deletion of edges not
nodes. The behaviour of
Hbatch
(
s2
) for
b >
1
.
4 indicates therefore an undesired
defect in the graph’s topology. Here the graph is split into numerous smaller
3.3. DATA-DRIVEN PATHWAYS 41
012
1000
2000
3000
b
no. of genes
012
0
200
400
600
800
b
no. of edges
0 1 2
0
200
400
600
800
b
no. of nodes
012
0.2
0.4
0.6
0.8
b
Edges/Nodes
t1
t2
t3
t4
t5
t6
t7
t8
Figure 3.1: Complete scan of
b
through all times (
t1, ..., t8
) of the stimulated batch
experiment, Gbatchpulse.
Upper left: up-regulated genes
Upper right: activated reactions (edges)
Lower left: activated metabolites (nodes)
Lower right: edges to nodes fraction
sub-graphs that are not connected.
With the same reasoning a sudden steep descent as present in figure 3.1 for
t7
shows a second defect, here the number of nodes and edges is approaching zero
because at b > 1.5 most nodes are being deleted.
Choice of b- Discrimination by Levels and Descent
The steepness of descent of the E/N can be regarded as uncertainty of the bound-
ary parameter
b
- in the close proximity of a chosen value the graphs topology
should not undergo strong alternations. For this reason, an analysis of the graph
should be carried out at some flat point of the curve.
In
Gbatch
E/N has a similar descent during both states until
b
= 1
.
5, then the
defect occurs. The same behaviour without defect for large values is visible for
Gconti
, here the descent is weaker for the state
s1
. Both graphs start at the
same level of E/N. It can be expected that the topology properties of the graphs
are similar too. Considering the stimulated batch culture, here both states are
immanently different, the short-term response is 0.2 smaller than the long-term
response, also their descents differ between each other: While it is strong for low
b
and then decreases close to zero for
s1
, it is constant for
s2
. The steeper descent
accounts for randomly spread edges. A constant level accounts for the deletion of
the same number of edges and nodes, hence nodes from the networks periphery
are being deleted for increasing b.
42 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
0 0.5 1 1.5 2
0.2
0.4
0.6
0.8
batch culture
b
E/N
G(s1)
G(s2)
H(s1)
H(s2)
0 0.5 1 1.5 2
0.2
0.4
0.6
0.8
continuous culture
b
E/N
0 0.5 1 1.5 2
0.2
0.4
0.6
0.8
stimulated batch culture
b
E/N
Figure 3.2: Scan of
b
for different transcriptome experiments. E/N is the fraction
of edges to nodes.
In order to study the effects of the choice of
b
, two different values,
b1= 0.8, b2= 1.3
,
will be compared for the batch culture. Since the stimulated batch culture is
referenced to an unstimulated batch culture, it is reasonable to choose
b
= 1
.
1 as
intermediate value of this interval.
3.3.2 Augmentation Characterises Solventogenesis
Acidogenesis and solventogenesis are two distinct states that are expected to be
visible in these networks in either experiment, batch culture or continuous culture.
For both it does not matter whether the graph is augmented during state
s2
, since
not much additional information is gained with respect to E/N. However, there
is a difference up to 0
.
2 between
G
(
s1
) and
H
(
s1
) in both experiments. This
eventually shows that a number of unconsidered reactions in
s1
are of no need in
s2
anymore and are therefore shut down. This indicates that solventogenic phase
is an adaptation to more hostile conditions. While the main metabolism from
s1
is largely preserved, additional reactions ensure the survival of the organism.
This becomes obvious in the continuous culture,
Hconti
(
s2
) is much larger than
Hconti
(
s1
) (table 3.2) and their intersection
Hconti,cX
is large, it contains one
fourth to one third of both graphs. In the converse
Hbatch
(
s2
) is smaller by 0.2
than
Hconti
(
s2
), here metabolic activity ceased because of starting sporulation.
Reactions that are relevant for sporulation are not covered by the KEGG database.
3.3. DATA-DRIVEN PATHWAYS 43
0 2 4 6
0
1
2
3
4
5
6
7
node degree distribution
node degree
number of nodes
data
fit
0 0.01 0.02 0.03 0.04 0.05
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
centrality measures - zoom
betweenness
closeness
Figure 3.3: Left: Node degree distribution in
GKEGG
CAC
with the linear fit for
determination of γ.
Right: Centrality statistics in
GKEGG
CAC
. Further centrality measures are not shown
as betweenness and stress are strongly correlated, as well as closeness, eccentricity
and radiality are strongly correlated.
3.3.3 Visualisation of the Standard Batch Fermentation
In the standard batch fermentation [
Jones et al., 2008
] in complex medium
s1
corresponds to
t
= 10
h
(figure 3.4) when solvent production starts, and
s2
corresponds to
t
= 40
h
(figure 3.5) when the culture enters the sporulation-state.
The reference state is the average over all transcripts and measured time-points.
The authors distinguish in their paper, six different clostridial stages occurring
in the temporal transcript expression data. The here chosen states are well
distinguishable according to these stages. Visualisation is focussed on revealing
the outcomes of two different boundary parameters.
Early Phase
The reactions (
b
= 1
.
3) during
s1
are sulfur aminoacid and serine metabolism,
co-factor synthesis, parts of sugar metabolism yielding butanoic acid and parts of
cell wall synthesis.
This view is complemented with the more uncertain reactions (
b
= 0
.
8), in partic-
ular for leucine- and gluthatione-synthesis. The pathway for mureine synthesis
44 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
Glycolate
2-Phosphoglycolate
L-Cysteine
Phosphoenolpyruvate
O-Phospho-4-hydroxy-L-threonine
4-Hydroxy-L-threonine
3-Oxododecanoyl-[acp]
Acetoacetyl-[acp]
Acyl-carrier
protein
3-Oxooctanoyl-[acp]
Hexanoyl-[acp]
Dodecanoyl-[acyl-carrier
protein]
Tetradecanoyl-[acp]
3-Oxotetradecanoyl-[acp]
Decanoyl-[acp]
CO2
Malonyl-[acp]
methyl ester
3-Oxodecanoyl-[acp]
3-Oxohexanoyl-[acp]
Acyl-[acyl-carrier
protein]
Butyryl-[acp]
3-Ketoglutaryl-[acp]
methyl ester
Glutaryl-[acp]
methyl ester
Malonyl-[acyl-carrier
protein]
Octanoyl-[acp]
3-Oxohexadecanoyl-[acp]
UDP-N-acetylmuramoyl-L-alanyl-D-glutamyl-6-carboxy-L-lysyl-D-alanyl-D-alanine
UMP
UDPMurAc(oyl-L-Ala-D-gamma-Glu-L-Lys-D-Ala-D-Ala)
MurAc(oyl-L-Ala-D-gamma-Glu-L-Lys-D-Ala-D-Ala)-diphospho-undecaprenol
Undecaprenyl-diphospho-N-acetylmuramoyl-L-alanyl-D-glutamyl-meso-2,6-diaminopimeloyl-D-alanyl-D-alanine
Acetyl-[acyl-carrier
protein]
di-trans,poly-cis-Undecaprenyl
phosphate
(S)-Malate
O-Succinyl-L-homoserine
Pyruvate
4-Phospho-L-aspartate
L-3,4-Dihydroxybutan-2-one
4-phosphate
L-Homoserine
L-2,3-Dihydrodipicolinate
Oxaloacetate
Orthophosphate
L-Aspartate
4-semialdehyde
di-trans,poly-cis-Undecaprenyl
diphosphate
2,3-Dehydroacyl-CoA
Riboflavin N3-Acetylgentamicin
C
Gentamicin C
N3'-Acetyl-2-deoxystreptamine
antibiotic
5-Amino-6-(1-D-ribitylamino)uracil
2-Deoxystreptamine
antibiotic
(S)-Lactate
Glyoxylate
D-4-Hydroxy-2-oxoglutarate
4-Hydroxy-2-oxoglutarate
6,7-Dimethyl-8-(D-ribityl)lumazine
Acetyl-CoA
Succinyl-CoA
3-(4-Hydroxyphenyl)pyruvate
Hexadecanoyl-[acp]
3-Oxostearoyl-[acp]
Prephenate
3-Ketopimeloyl-[acp]
methyl ester
Chorismate
L-Threonine
NADH
ADP
L-Glutamyl
5-phosphate
5-Phosphoribosylamine
dADP
2-Butenoate
2-Amino-7,8-dihydro-4-hydroxy-6-(diphosphooxymethyl)pteridine
NAD+
beta-D-Fructose
Acetaldehyde Glycine
L-Allothreonine
2-Dehydro-3-deoxy-D-arabino-heptonate
7-phosphate
2-Dehydro-3-deoxy-D-gluconate
AMP
Butanoylphosphate
D-Ribose
5-phosphate
ADP-ribose
O-Acetyl-L-serine
NH3
2-Dehydro-3-deoxy-6-phospho-D-gluconate
D-Glyceraldehyde
3-phosphate
Formaldehyde
5,10-Methylenetetrahydrofolate
(R)-Pantoate
IMP
2-Dehydropantoate
D-Glucosamine
6-phosphate
beta-D-Fructose
6-phosphate
3-Phosphonooxypyruvate
O-Phospho-L-homoserine
Butanoic acid
ATP
D-Fructose
6-phosphate
3-Phospho-D-glycerate
D-Fructose
(S)-3-Methyl-2-oxopentanoic
acid
2-Oxoglutarate 4-Methyl-2-oxopentanoate
3-Methyl-2-oxobutanoic
acid 2-Hydroxyglutarate
L-Isoleucine
L-Leucine
L-Valine
L-Glutamine
Deamino-NAD+
L-Glutamate
5'-Phosphoribosylglycinamide
L-Asparaginyl-tRNA(Asn)
D-Galactosyl-N-acetyl-D-galactosaminyl-(N-acetylneuraminyl)-D-galactosyl-D-glucosylceramide
N-Acetyl-D-galactosaminyl-(N-acetylneuraminyl)-D-galactosyl-D-glucosylceramide
L-Leucyl-tRNA
Galactan
GA2
beta-D-Glucosyl-(1<->1)-ceramide
beta-D-Galactosyl-(1->4)-beta-D-glucosyl-(1<->1)-ceramide
D-Galactose
tRNA(Leu)
GA1
Protein histidine
alpha-D-Glucose
Lactose
alpha-D-Glucose
6-phosphate
Protein
N(pi)-phospho-L-histidine
D-Glucose
Lactose
6-phosphate
Maltose
D-Fructose
1-phosphate
alpha,alpha-Trehalose
alpha,alpha'-Trehalose
6-phosphate
Maltose
6'-phosphate
N-Acetylmuramic
acid 6-phosphate
Sucrose
6-phosphate
D-Mannose
6-phosphate
Salicin
6-phosphate
Salicin
Galactitol
1-phosphate
D-Mannose
N-Acetylmuramate
D-Galactosamine
Sucrose
Galactitol
L-Sorbose
D-Sorbitol
D-Mannitol
1-phosphate
Mannitol
L-Ascorbate
6-phosphate
Arbutin
Sorbitol
6-phosphate
Sorbose
1-phosphate
Ascorbate
Arbutin
6-phosphate
D-Glucosamine
N-Acetyl-D-glucosamine
6-phosphate
5-Phospho-alpha-D-ribose
1-diphosphate
N-Acetyl-D-galactosamine
6-phosphate
N-Acetyl-D-galactosamine
N-Acetyl-D-glucosamine
D-Galactosamine
6-phosphate
Penicilloic acid(3S)-3-Hydroxyadipyl-CoA(15S)-15-Hydroxy-5,8,11-cis-13-trans-eicosatetraenoate5(S)-HPETE
5(S)-HETE CellobiosetRNA uridine15(S)-HPETE Ribonucleoside
diphosphate
3-Oxoadipyl-CoA Penicillin MercaptopyruvatePenicillin G
UDP-N-acetyl-alpha-D-glucosamine
dUDP
UDP-N-acetyl-D-mannosamine
UDP
N-Acetyl-D-mannosamine
1-(5'-Phosphoribosyl)-5-amino-4-(N-succinocarboxamide)-imidazole
L-Aspartate
beta-D-Glucose
1-(5-Phospho-D-ribosyl)-5-amino-4-imidazolecarboxylate
3-Hydroxybutanoyl-CoA
(S)-3-Hydroxybutanoyl-CoA
5-Carboxyamino-1-(5-phospho-D-ribosyl)imidazole
3-Keto-beta-D-galactose
tRNA
pseudouridine dGDPPhenylpropanoate Thioredoxin
disulfide NADP+dCDPCellulose3-Mercaptolactate
2'-Deoxyribonucleoside
diphosphate
Benzylpenicilloic
acid
GDP ThioredoxinCDPtrans-Cinnamate NADPH Hydrogen
peroxide
meso-2,6-Diaminoheptanedioate
D-Glyceraldehyde
L-Fucose
1-phosphate
Thiol-containing
reductant
Trypanothione
disulfide
L-Ornithine
UreaUDP-N-acetylmuramoyl-L-alanyl-D-glutamate
UDP-N-acetylmuramoyl-L-alanyl-D-gamma-glutamyl-meso-2,6-diaminopimelate D-GlycerateL-ArginineAcetoacetyl-CoA
Oxidized
thiol-containing
reductant
6-Deoxy-L-galactose Trypanothione3-Ketolactose
tRNA(Ser)
Phosphatidate
L-Seryl-tRNA(Sec)
1-Acyl-sn-glycerol
3-phosphate
tRNA(Sec)
Phosphatidylserine
L-Seryl-tRNA(Ser)
Glycerophosphodiester
CDP-diacylglycerol
Choline
Ethanolamine sn-glycero-3-Phosphocholine
sn-Glycerol
3-phosphate
Phosphatidylglycerophosphate
CMP
4-(Cytidine
5'-diphospho)-2-C-methyl-D-erythritol
2-C-Methyl-D-erythritol
4-phosphate
sn-glycero-3-Phosphoethanolamine
D-Ribulose
5-phosphate
Butanoyl-CoA
Formate
CoA
Propanoyl-CoA
2-Hydroxybutanoic
acid
2-Oxobutanoate
L-Serine
Acyl-CoA L-Methionine
Alcohol
S-Adenosyl-L-homocysteine
S-Adenosyl-L-methionine
Methanethiol
5'-Deoxyadenosine
5-Amino-6-(5'-phosphoribosylamino)uracil
O-Acetyl-L-homoserine
CTP
Xanthine
dUMP
dUTP
Nicotinate
Methylselenic acid
Methaneselenol
5-Amino-6-(5'-phospho-D-ribitylamino)uracil
Se-Methyl-L-selenocysteine
Acetate
Diphosphate
2,5-Diamino-6-(5-phospho-D-ribosylamino)pyrimidin-4(3H)-one
GTP
Nicotinate
D-ribonucleotide
GMP
Guanine
Xanthosine
5'-phosphate
Hydrogen sulfide
L-Homocysteine
ROH
Phospholipid
olefinic fatty acid
ROOH
Phospholipid
cyclopropane
fatty acid
2-(Formamido)-N1-(5'-phosphoribosyl)acetamidine
Glutaminyl-tRNA
Aminoimidazole
ribotide
L-Glutamyl-tRNA(Gln)
5,10-Methenyltetrahydrofolate
Tetrahydrofolate
L-Aspartyl-tRNA(Asn)
10-Formyltetrahydrofolate
5'-Phosphoribosyl-N-formylglycinamide
1-(5'-Phosphoribosyl)-5-amino-4-imidazolecarboxamide
1-(5'-Phosphoribosyl)-5-formamido-4-imidazolecarboxamide
Glycolaldehyde
2-Amino-4-hydroxy-6-(D-erythro-1,2,3-trihydroxypropyl)-7,8-dihydropteridine
2-Amino-4-hydroxy-6-hydroxymethyl-7,8-dihydropteridine
Selenite
Hydrogen
selenide
D-Erythrose
4-phosphate
L-Selenocysteine
L-Alanine
[Enzyme]-S-sulfanylcysteine
S-Glutathionyl-L-cysteine
Sulfite
[Enzyme]-cysteine
Glutathione
Glutathione
disulfide
Thiosulfate
Figure 3.4: Early Batch Experiment (
Hbatch
(
s1
)), darkblue and lightblue:
b >
0
.
8;
darkblue: b > 1.3
is added. A delta-2-oxidreductase for crotonic acid becomes apparent. Also a
threonine lyase is active.
From the graph it is easily visible, that lowering
b
does not interlink existing
compounds but instead, it adds new branches to the existing network.
Late Phase
Compared to the early phase, the late phase contains less reactions and metabolites
for the stronger certitude than for the weaker - here, one obtain the same number of
metabolites which are more densely connected. For
b
= 1
.
3 the energy metabolism,
3.3. DATA-DRIVEN PATHWAYS 45
N-Acetylmuramoyl-Ala
N-Acetylmuramate
(R)-Lactate
N-Acetyl-D-glucosamine
D-Mannose
6-phosphate
Mannitol
L-Ascorbate
6-phosphate Sorbose
1-phosphate
D-Galactosamine
L-Sorbose
Ascorbate
D-Glucosamine
D-Mannose
L-Alanine
D-Galactosamine
6-phosphate
N-Acetylmuramic
acid 6-phosphate
N-Acetyl-D-glucosamine
6-phosphate
D-Glucosamine
6-phosphate
Molybdoenzyme
molybdenum
cofactor
Adenylated
molybdopterin
Glycerone
phosphate
D-Erythrose
4-phosphate
L-Glutamate
5-semialdehyde
(R)-2,3-Dihydroxy-3-methylpentanoate
Molybdate
(R)-2,3-Dihydroxy-3-methylbutanoate
D-Tagatose
6-phosphate beta-D-Glucose
1-phosphate
2-C-Methyl-D-erythritol
2,4-cyclodiphosphate
2-Phospho-4-(cytidine
5'-diphospho)-2-C-methyl-D-erythritol
D-Glyceraldehyde
Phosphoenolpyruvate
D-Fructose
1-phosphate D-Glucose
6-phosphate
2-(alpha-Hydroxyethyl)thiamine
diphosphate
Phosphatidylserine
Thiamin
diphosphate
(R)-3-Hydroxy-3-methyl-2-oxopentanoate
CDP-diacylglycerol
CMP
(S)-2-Acetolactate
3-Hydroxy-3-methyl-2-oxobutanoic
acid
2-Acetolactate
2,3-Dihydroxy-3-methylbutanoate
L-Aspartate
4-semialdehyde
alpha-D-Glucose
6-phosphate
alpha-D-Glucose
Sucrose
beta-D-Fructose
beta-D-Glucose
6-phosphate
ROH D-Glucoside
alpha,beta-Dihydroxyethyl-TPP
D-Xylulose
5-phosphate
Sedoheptulose
7-phosphate
D-Fructose
6-phosphate
D-arabino-Hex-3-ulose
6-phosphate
Aminofructose
6-phosphate
D-Ribose
5-phosphate
D-Ribulose
5-phosphate
Formaldehyde
Iminoerythrose
4-phosphate
beta-D-Fructose
6-phosphate
Pyruvate
L-Serine
(S)-2-Aceto-2-hydroxybutanoate
1-Deoxy-D-xylulose
5-phosphate
L-2,3-Dihydrodipicolinate
beta-D-Fructose
1,6-bisphosphate
Sedoheptulose
1,7-bisphosphate
D-Fructose
1,6-bisphosphate
D-Glyceraldehyde
3-phosphate
D-Fructose
2-Phospho-D-glycerate
Maltose
6-Phospho-beta-D-glucosyl-(1,4)-D-glucose
Alcohol 6-Phospho-beta-D-galactoside D-Galactose
6-phosphate
Precorrin 4
L-Methionine
(S)-S-oxide
Precorrin 5
L-Methionine
S-Adenosyl-L-homocysteine
5'-Deoxyadenosine
S-Adenosyl-L-methionine
S-Adenosyl-4-methylthio-2-oxobutanoate
4-Methyl-2-oxopentanoate
Biotin
(2S)-2-Isopropyl-3-oxosuccinate
Dethiobiotin
7,8-Diaminononanoate
CO2
2-Methylmaleate
D-erythro-3-Methylmalate
N-(5-Phospho-D-ribosyl)anthranilate
Isopentenyl
diphosphate
N-Carbamoyl-L-aspartate
5-Phospho-alpha-D-ribose
1-diphosphate
1-Hydroxy-2-methyl-2-butenyl
4-diphosphate
Dimethylallyl
diphosphate
dATP
Diphosphate
1-(5-Phospho-D-ribosyl)-ATP
RNA
L-Tryptophan
2-Oxobutanoate
Indole
Indoleglycerol
phosphate (R)-2-Methylmalate Cobalt-precorrin 4
Cobalt-precorrin
5A
8-Amino-7-oxononanoate
10-Formyltetrahydrofolate L-Methionyl-tRNA
Tetrahydrofolate
Chloroacetic acid Glycolate Glyoxylate
Hydrochloric acid
Orthophosphate
N-Formylmethionyl-tRNA
ATP
Carbamoyl
phosphate
2-Deoxy-D-ribose
5-phosphate
dTMP
dUMP
5,10-Methylenetetrahydrofolate
Dihydrofolate
1-(2-Carboxyphenylamino)-1-deoxy-D-ribulose
5-phosphate
AMP
dAMP
Shikimate
DNA
Phosphatidylethanolamine
Choline
phosphate
Ethanolamine
phosphate
Phosphatidylcholine
O-1-Alk-1-enyl-2-acyl-sn-glycero-3-phosphoethanolamine
1-Alkenyl-2-acylglycerol
UMP
Uracil
3-Phosphonooxypyruvate
Propanoate
D-Glycerate
3-Phospho-D-glycerate
dGMP
dGDP
Propanoyl
phosphate
L-Glutamyl
5-phosphate
D-Tagatose
1,6-bisphosphate
D-Glucose
1-phosphate
Starch
Acetate
Shikimate
3-phosphate
dADP
Acetyl phosphate
GDP
ADP
GMP
ADP-glucose
Deoxynucleoside
triphosphate
dGTP
GTP
Nucleoside
triphosphate
dTTP
UTP
dCTP
CTP
1-Phosphatidyl-D-myo-inositol
sn-Glycerol
3-phosphate
Phosphatidylglycerol
1,2-Diacyl-sn-glycerol
Inositol
1-phosphate
Holo-[carboxylase]
Carboxybiotin-carboxyl-carrier
protein
HCO3-
Malonyl-CoA
2-Methylacetoacetyl-CoA
Propanoyl-CoA
N-Acetyl-L-glutamate
L-Ornithine
Butanal
Butanoic acid
3-Oxohexanoyl-CoA
Citrate
cis-Aconitate
Acetoacetyl-CoA
Acetyl-CoA
Ethanol
(2R,3S)-3-Isopropylmalate
Acetone
Acetoacetate
2-Isopropylmaleate
alpha-Isopropylmalate
3-Methyl-2-oxobutanoic
acid
Isocitrate
Acetaldehyde
Butanoyl-CoA
Chorismate
Ethanolamine
NH3
L-Glutamine
alpha-Amino acid
Hydrazine
Nitrogen
Glutaminyl-tRNA
Diimine
Succinate
semialdehyde
3-Oxopropanoate
3-Sulfopyruvate
3-Sulfinylpyruvate
L-Cysteine
L-Tyrosine
4-Aminobutanoate
beta-Alanine
gamma-L-Glutamyl-L-cysteine
L-Cysteate
3-Sulfino-L-alanine
L-Phenylalanine
Phenylpyruvate
L-Aspartate
Oxaloacetate
Dephospho-CoA
L-erythro-4-Hydroxyglutamate
D-4-Hydroxy-2-oxoglutarate
Mercaptopyruvate
L-Glutamate
DL-Glutamate
3-(4-Hydroxyphenyl)pyruvate
CoA
Succinyl-CoA
Anthranilate
2-Oxoglutarate
2-Hydroxyglutarate
L-Glutamyl-tRNA(Gln)
N-Acetylornithine
L-Asparaginyl-tRNA(Asn)
L-Aspartyl-tRNA(Asn)
3-Ketoglutaryl-[acp]
methyl ester
3-Ketopimeloyl-[acp]
methyl ester Aldophosphamide
2-Phenyl-1,3-propanediol
monocarbamate
Choline6-Deoxy-L-galactose 4-Hydroxy-5-phenyltetrahydro-1,3-oxazin-2-one
2',3'-Cyclic GMP
Phosphoenol-4-deoxy-3-tetrulosonate Peptide-L-methionine
(S)-S-oxide
trans-3-Chloroallyl
aldehyde
Peptide-L-methionine3-Phospho-D-erythronate
cis-3-Chloroallyl
aldehyde
Guanosine
3'-phosphate
(R)-3-Hydroxydodecanoyl-[acp]
3-Oxododecanoyl-[acp]trans-3-Chloro-2-propene-1-olcis-3-Chloro-2-propene-1-ol
(S)-2-Haloacid
Halide L-Histidinol
(R)-2-Hydroxyacid
L-Histidine
6-Pyruvoyltetrahydropterin
Thioredoxin
disulfide 2',3'-Cyclic UMP
3-(Imidazol-4-yl)-2-oxopropyl
phosphate
D-erythro-1-(Imidazol-4-yl)glycerol
3-phosphate
3'-AMP
2',3'-Cyclic AMP
3'-UMP2',3'-Cyclic CMPThioredoxin 3'-CMP
L-Histidinal 3-Carbamoyl-2-phenylpropionaldehyde5-Phenyl-1,3-oxazinane-2,4-dioneBetaine aldehyde
L-Fucose
1-phosphate
3-Hydroxyglutaryl-[acp]
methyl ester Alcophosphamide
3-Hydroxypimeloyl-[acp]
methyl ester
3,4-Dihydroxyphenylethyleneglycol3-Oxohexanoyl-[acp] (3R)-3-Hydroxytetradecanoyl-[acyl-carrier
protein]
3alpha,7alpha,26-Trihydroxy-5beta-cholestane
3alpha,7alpha-Dihydroxy-5beta-cholestan-26-al3-Oxotetradecanoyl-[acp]3,4-Dihydroxymandelaldehyde(R)-3-Hydroxyhexanoyl-[acp]
Retinal D-Tagaturonate D-Galacturonate
5-Dehydro-D-gluconate
D-Gluconic acid D-Fructuronate D-Glucuronate
Retinol Xanthosine
5'-phosphate
Ketone
Secondary alcohol
2-Oxo acid IMP
(S)-2-Hydroxyacid
alpha,alpha'-Trehalose
6-phosphate
alpha,alpha-Trehalose
Galactitol
1-phosphate
D-Sorbitol
N-Acetyl-D-galactosamine
Sorbitol
6-phosphate
Salicin
6-phosphate
N-Acetyl-D-galactosamine
6-phosphate Protein
N(pi)-phospho-L-histidine
D-Glucose
Maltose
6'-phosphate
Lactose
6-phosphate
Arbutin
6-phosphate
Cyanoglycoside
Prunasin
cis-2-Hydroxycinnamate
cis-beta-D-Glucosyl-2-hydroxycinnamate
Cellobiose
Amygdalin
Lactose
Cellulose
D-Alanine
Salicyl alcohol
Mandelonitrile
Protein histidine
D-Mannitol
1-phosphate
Galactitol
Salicin
Cyanohydrin
Linamarin
beta-D-Glucose
Acetone
cyanohydrin
Sucrose
6-phosphate
Hydroquinone
Arbutin
(2R)-2-Hydroxy-2-methylbutanenitrile
(S)-4-Hydroxymandelonitrile
Lotaustralin
Dhurrin
(3R)-3-Hydroxybutanoyl-[acyl-carrier
protein]
(3R)-3-Hydroxydecanoyl-[acyl-carrier
protein]
3-Oxodecanoyl-[acp]
(3R)-3-Hydroxyoctanoyl-[acyl-carrier
protein]
3-Oxooctanoyl-[acp]3-Oxohexadecanoyl-[acp] Acetoacetyl-[acp]
(3R)-3-Hydroxypalmitoyl-[acyl-carrier
protein]
7,8-Dihydroneopterin
3'-triphosphate
TriphosphatePhosphoribosyl-AMP
Formyl-L-methionyl
peptide
N-(5'-Phospho-D-1'-ribulosylformimino)-5-amino-1-(5''-phospho-D-ribosyl)-4-imidazolecarboxamide
Formate
5-(5-Phospho-D-ribosylaminoformimino)-1-(5-phosphoribosyl)-imidazole-4-carboxamide
Methionyl peptide
NAD+
UDP-alpha-D-galactose NADH
UDP-N-acetyl-alpha-D-glucosamine NADPH
NADP+
UDP-glucose
L-Arabinose
2,4-Diamino-6-hydroxylaminotoluene3-Hydroxyoctadecanoyl-[acp] Trichloroethanol 2-Naphthaldehyde 1-Naphthaldehyde Ethylene6-Thioinosine-5'-monophosphate
(3R)-3-Hydroxyacyl-[acyl-carrier
protein]
3-Oxoacyl-[acyl-carrier
protein]
2',3'-Cyclic
nucleotide dTDP-galactose dTDP-glucose 1-Octanal 1-Octanol
Nucleoside
3'-phosphate
Acetylene1-Hydroxymethylnaphthalene2,4-Diamino-6-nitrotoluene (2-Naphthyl)methanolChloral hydrate3-Oxostearoyl-[acp]
6-Thioxanthine
5'-monophosphate
OxygenUDP-N-acetyl-D-galactosamineAldehyde Hydrogen
peroxide
Primary alcohol
Figure 3.5: Late Batch Experiment (
Hbatch
(
s2
)), darkblue and lightblue
b >
0
.
8 ,
darkblue: b > 1.3
glycolytic paths, pentose pathways and secondary metabolism become apparent,
aspartate metabolism and glutamate are central. On a less certain level nucleotide
synthesis, butanal production and membrane biosynthesis are seen. The reactions
for
b
= 0
.
8 are connections of the more certain reactions for
b
= 1
.
3: Acetate,
O-acetyl-serine and acetyl-CoA are the most stressed metabolites (not shown)
for
b
= 0
.
8. As expected, they play a central role during
s1
. For
b
= 1
.
3, these
metabolites move back in ranking to places 8, 5 and 6 respectively.
46 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
dUTP
L-Glutamine
L-Glutamate
dCTP CTP
UTP
5,6,7,8-Tetrahydromethanopterin
NH3
5-Methyl-5,6,7,8-tetrahydromethanopterin
Methyl-Co(III)
corrinoid protein
HCO3- Carbamoyl
phosphate
Co(I) corrinoid
protein
dGDP
L-Aspartate
N-Carbamoyl-L-aspartate
dCDP GDP
dADP
CDP
ATP ADP
CO
CO2
2-Oxoglutarate
CoA
Succinyl-CoA
Acetyl-CoA
Orthophosphate Succinate NADP+ NADPH
Fumarate
UDP
dUDP
(15S)-15-Hydroxy-5,8,11-cis-13-trans-eicosatetraenoate
Peptide-L-methionine
(R)-S-oxide
Sedoheptulose
1,7-bisphosphate
UMP
D-Fructose
1,6-bisphosphate
D-Fructose
6-phosphate
beta-D-Fructose
1,6-bisphosphate
15(S)-HPETEPeptide-L-methionine
5(S)-HETE
(S)-Dihydroorotate
Thioredoxin
Thioredoxin
disulfide Orotate
5(S)-HPETE
L-Cysteine
Ribonucleoside
diphosphate
S-Glutathionyl-L-cysteine
beta-D-Fructose
6-phosphate
Glutathione
Sedoheptulose
7-phosphate Glutathione
disulfide
2'-Deoxyribonucleoside
diphosphate
Hydrogen
selenide
Trypanothione
disulfide
Methylselenic acid
Selenite
2,4-Diamino-6-hydroxylaminotoluene
D-Galactose
6-phosphate
Orotidine
5'-phosphate
D-Tagatose
6-phosphate
2,4-Diamino-6-nitrotoluene
TrypanothioneMethaneselenol
Figure 3.6: Early Batch Pulse Experiment (
Gbatchpulse
(
s1
)), colour:green to red
for decreasing eccentricity, size: small to large for increasing stress
3.3.4 Visualisation of the Acetic Acid Pulse Experiment in Batch
In the acetic acid pulse experiment [
Alsaker et al., 2010
],
s1
corresponds to the
short-term response after 0
.
25
h
(figure 3.6) and
s2
to the long-term response after
20
h
(figure 3.7). Here, the reference state is an independent batch culture and for
each measured time-point, an untreated reference is used. Therefor, one must not
augment the
G
-graph. Visualisation is focussed on the different graph topological
parameters stress and eccentricity and their role to increase human-readability of
the network.
Short-term Response
For the short-term response, the network is small - only few paths are activated
after the pulse. Notably, reactions around ammonia and phosphate become
apparent, conversion of fructose-bis-phosphate and seduheptulose phosphates,
3.3. DATA-DRIVEN PATHWAYS 47
glutamine and aspartate conversion. Further, few reactions are involving acetyl-
CoA. Thioredoxin utilization increases by multiple reactions.
Long-term Response
The long-term response is more complex than the short-term response. It involves
a variety of CoA-reactions, a multitude of carbon dioxide involving reactions.
here are butanoyl-CoA originating reactions, synthesis of different amino acids
and acetyl-CoA driven synthesis of branched small fatty acids. Further, the
ABC-transporters for sugars are activated. One also finds upgregulated sugar
import in the batch culture, this suggests that acetic acid has a stimulating effect
on glucose uptake. This was seen also in continuous culture (data not shown).
From a comparison with
Hconti
(
s2
), one recognises that more than two third of
the involved reactions are identical to the late stimulated network. Acetic acid
addition seems indeed be an inducer for a state comparable to solventogenesis,
this also was seen in continuous culture, acetone and butanol were produced after
short and sustained stimuli with acetic acid (data not shown).
3.3.5 Visualisation of the pH-Shift Experiment in Continuous Culture
The pH-shift experiment [
Grimmler et al., 2011
] is more useful with respect to the
batch fermentation experiment because of two aspects: The two phases are clearly
separated before and after the shift and no sporulation occurs in the continuous
culture. Acidogenic conditions correspond to
s1
, early solventogenic condition,
when the pH is stabilised to
s2
. Visualisation is focussed on the comparison of
augmented and non-augmented graph and centralities of the augmented graph.
The Graph of Acidogenesis
Graph centrality measures are shown in figure 3.8 and complemented with a
third dimension, the
G
- and
H
-graph (figure 3.9). Glutamate and joint amino
acids like glutamine and aspartate take a central role of this pathway. In a more
distant region different CoA and phosphate derivatives are present. The most
outside regions are occupied by several fatty acid metabolites and vitamines. Most
interestingly for this network in contrast to solventogenesis, the carbon dioxide
node is only loosely connected to the overall network.
There is the carbon monoxide fixation pathway, known to be active for several
organisms and Clostridia, but not in C. acetobutylicum [
Koepke et al., 2011
], the
decarboxylation of isocitrate to oxoglutarate and the oxidative decarboxylation of
oxoglutarate to succinyl-CoA. All these reactions belong to the H-graph, hence
are strongly down-regulated in solventogenesis. There is no other decarboxylase
up-regulated in acidogenesis.
In addition, the whole conversion path from acetyl-CoA to butanal is already
48 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
Ubiquinol
3-Hydroxypimeloyl-[acp]
methyl ester 4-Hydroxy-5-phenyltetrahydro-1,3-oxazin-2-oneD-Glycerate6-Deoxy-L-galactose
3-Hydroxyglutaryl-[acp]
methyl ester Choline Trypanothione
Ubiquinone
L-Fucose
1-phosphate Betaine aldehyde
Enoylpimeloyl-[acp]
methyl ester 5-Phenyl-1,3-oxazinane-2,4-dione
Precorrin 6Y
D-Glyceraldehyde
Enoylglutaryl-[acp]
methyl ester
Trypanothione
disulfide
Isocitrate
L-Histidinal
L-HistidineCitrate
L-Histidinol
cis-Aconitate
6-Pyruvoyltetrahydropterin
Peptide-L-methionine
(S)-S-oxide Triphosphate
Peptide-L-methionine
(R)-S-oxide
Peptide-L-methionine 7,8-Dihydroneopterin
3'-triphosphate
(S)-3-Methyl-2-oxopentanoic
acid
(R)-2,3-Dihydroxy-3-methylpentanoate
cis-3-Chloroallyl
aldehyde
trans-Dodec-2-enoyl-[acp]
1-Naphthaldehyde
Cobalt-dihydro-precorrin
6
(R)-3-Hydroxydodecanoyl-[acp]
trans-3-Chloroallyl
aldehyde
trans-3-Chloro-2-propene-1-olcis-3-Chloro-2-propene-1-ol1-Hydroxymethylnaphthalene
2'-Deoxyribonucleoside
diphosphate
1-Octanol
2-Deoxy-D-ribose
5-phosphate
2-Deoxy-D-ribose
1-phosphate
Mercaptopyruvate
1-Octanal
Ribonucleoside
diphosphate
3-Mercaptolactate
Precorrin 8X
Cobalt-precorrin 8
But-2-enoyl-[acyl-carrier
protein]
(3R)-3-Hydroxybutanoyl-[acyl-carrier
protein]
Cobalt-precorrin 7
2-Phenyl-1,3-propanediol
monocarbamate
D-Glucose
3-Carbamoyl-2-phenylpropionaldehyde
Alcophosphamide
D-Sorbitol
N-Acetylmuramate
Lactose
alpha,alpha-Trehalose
Phosphoenolpyruvate
1-(5-Phospho-D-ribosyl)-ATP
Shikimate
L-Aspartate
4-semialdehyde
4-Phospho-L-aspartate
Diphosphate
Orthophosphate
Shikimate
3-phosphate
Fe2+
GTP
Fe3+
NAD+
NADH
5-Amino-6-(5'-phosphoribosylamino)uracil
3-Phospho-D-glyceroyl
phosphate
alpha-D-Ribose
1-phosphate
5-Amino-6-(1-D-ribitylamino)uracil
IMP
Xanthosine
5'-phosphate
Riboflavin
D-Ribose
5-phosphate
Thioredoxin
disulfide
Thioredoxin
2,3-Bisphospho-D-glycerate
2,5-Diamino-6-(5-phospho-D-ribosylamino)pyrimidin-4(3H)-one
CDP
trans-Hex-2-enoyl-[acp]
dCDP
L-Aspartyl-tRNA(Asp)
tRNA(Asp)
Retinal
Retinol
5(S)-HETE15(S)-HPETE
Trichloroethanol
(2-Naphthyl)methanol
2-Naphthaldehyde
Chloral hydrate
(15S)-15-Hydroxy-5,8,11-cis-13-trans-eicosatetraenoate 5(S)-HPETE
6-Thioinosine-5'-monophosphate
(3R)-3-Hydroxypalmitoyl-[acyl-carrier
protein]
Aldophosphamide 6-Thioxanthine
5'-monophosphate
trans-Hexadec-2-enoyl-[acp]trans-Tetradec-2-enoyl-[acp]
(3R)-3-Hydroxytetradecanoyl-[acyl-carrier
protein]
5-Amino-6-(5'-phospho-D-ribitylamino)uracil
(3R)-3-Hydroxydecanoyl-[acyl-carrier
protein]
(3R)-3-Hydroxyoctanoyl-[acyl-carrier
protein]
trans-Dec-2-enoyl-[acp]trans-Oct-2-enoyl-[acp]
D-Ribulose
5-phosphate
Formate
L-3,4-Dihydroxybutan-2-one
4-phosphate
O-Phospho-4-hydroxy-L-threonine
4-Hydroxy-L-threonine
6,7-Dimethyl-8-(D-ribityl)lumazine
3-Oxodecanoyl-[acp]
3-Oxooctanoyl-[acp]
3-Oxohexanoyl-[acp]
Hexanoyl-[acp]
Octanoyl-[acp]
Butyryl-[acp]
Malonyl-[acp]
methyl ester
3-Ketoglutaryl-[acp]
methyl ester
Acyl-carrier
protein
Hexadecanoyl-[acp]
3-Oxostearoyl-[acp]
Glutaryl-[acp]
methyl ester
Decanoyl-[acp]
Dodecanoyl-[acyl-carrier
protein]
Tetradecanoyl-[acp]
3-Ketopimeloyl-[acp]
methyl ester
Acyl-[acyl-carrier
protein]
Acetyl-[acyl-carrier
protein]
2-Isopropylmaleate
Thiamin
diphosphate
L-Phenylalanine
Protoporphyrinogen
IX
2-Oxobutanoate
4-Methyl-2-oxopentanoate
D-erythro-3-Methylmalate
2-Dehydro-3-deoxy-6-phospho-D-gluconate
Acetoacetyl-[acp]
Secondary alcohol NADP+FADH2 Primary alcohol Hydrogen
peroxide
D-Glyceraldehyde
3-phosphate
L-Seryl-tRNA(Sec)
3-Phosphonooxypyruvate
tRNA(Sec)
Indole
2-Phospho-D-glycerate
HCO3-
3-Oxohexadecanoyl-[acp]
3-Oxotetradecanoyl-[acp]
Carbonic acid
L-Arogenate
Coproporphyrinogen
III
3-Oxododecanoyl-[acp]
Malonyl-[acyl-carrier
protein]
3-Phospho-D-glycerate
L-Serine
Indoleglycerol
phosphate
L-Tryptophan
L-Seryl-tRNA(Ser)
(R)-2-Methylmalate
tRNA(Ser)
(S)-2-Acetolactate
2-(alpha-Hydroxyethyl)thiamine
diphosphate
(2S)-2-Isopropyl-3-oxosuccinate
(S)-2-Aceto-2-hydroxybutanoate
2-Acetolactate
(2R,3S)-3-Isopropylmalate
2-Hydroxybutanoic
acid
O-Acetyl-L-serine
OxygenAldehydeFAD NADPHKetone
2-Methylmaleate
3,4-Dihydroxyphenylethyleneglycol
3,4-Dihydroxymandelaldehyde3alpha,7alpha-Dihydroxy-5beta-cholestan-26-al
dUDP
3alpha,7alpha,26-Trihydroxy-5beta-cholestane
Ascorbate
Selenite
L-Ascorbate
6-phosphate
Protein
N(pi)-phospho-L-histidine
Hydrogen
selenide
Galactitol
1-phosphate
N-Acetyl-D-glucosamine
6-phosphate
Arbutin
Salicin
Maltose
N-Acetyl-D-galactosamine
D-Mannose
Sucrose
D-Fructose
D-Fructose
6-phosphate
D-Mannitol
1-phosphate
D-Glucosamine
6-phosphate
Lactose
6-phosphate
Galactitol
Protein histidine
D-Glucosamine
Sorbose
1-phosphate
D-Galactosamine
6-phosphate
D-Mannose
6-phosphate
D-Galactosamine
Sucrose
6-phosphate
D-Fructose
1-phosphate
N-Acetylmuramic
acid 6-phosphate
alpha-D-Glucose
6-phosphate
Sorbitol
6-phosphate
Mannitol
alpha,alpha'-Trehalose
6-phosphate
L-Sorbose
Acetate
Salicin
6-phosphate
N-Acetyl-D-glucosamine
N-Acetyl-D-galactosamine
6-phosphate
Arbutin
6-phosphate
Maltose
6'-phosphate
L-Selenocysteine
3-Oxopropanoate
4-Aminobutanoate
beta-Alanine
2-Hydroxyglutarate
L-Glutamine
Succinyl-CoA
Adenosine
Succinate
semialdehyde
Methaneselenol
Adenine
Deoxyadenosine
(R)-1-Aminopropan-2-ol Dolichyl
phosphate
Dolichyl
phosphate
D-mannose
Adenosyl
cobinamide
Phosphatidate
1-Acyl-sn-glycerol
3-phosphate
Deoxyinosine
Inosine
Hypoxanthine
Methylselenic acid
Glutaminyl-tRNA
3-Methyl-2-oxobutanoic
acid
Butanoic acid
Butanal
Acetoacetate
2,3-Dihydroxy-3-methylbutanoate
CoA
Butanoyl-CoA
2,3-Dehydroacyl-CoA
(R)-2,3-Dihydroxy-3-methylbutanoate
Acyl-CoA
Acetoacetyl-CoA
alpha-Isopropylmalate
2-Methylacetoacetyl-CoA
Acetaldehyde
Ethanol
Acetyl-CoA
Propanoyl-CoA
3-Oxohexanoyl-CoA
[Enzyme]-S-sulfanylcysteine
Glutathione
disulfide
[Enzyme]-cysteine
S-Glutathionyl-L-cysteine
Glutathione
Adenosyl
cobyrinate
hexaamide
D-1-Aminopropan-2-ol
O-phosphate
Adenosyl
cobinamide
phosphate
L-Glutamyl-tRNA(Gln)
GDP
GDP-mannose
L-Cysteine
dGDP
L-Alanine
5-O-(1-Carboxyvinyl)-3-phosphoshikimate
5-Guanidino-2-oxopentanoate
D-Lysine
D-Arginine
5-Phospho-alpha-D-ribose
1-diphosphate
6-Amino-2-oxohexanoate
D-Amino acid
D-Ornithine
5-Amino-2-oxopentanoic
acid
Nicotinate
D-ribonucleotide
Quinolinate
2-Oxo acid
CO2
(S)-Lactate
Glyoxylate
D-Alanine
Prephenate
Se-Methyl-L-selenocysteine
L-erythro-4-Hydroxyglutamate
L-Glutamate
D-4-Hydroxy-2-oxoglutarate
alpha-Amino acid
Chorismate
Iminoaspartate
(R)-3-Hydroxyhexanoyl-[acp]
NH3
Anthranilate
N-(5-Phospho-D-ribosyl)anthranilate
L-Aspartate
Oxaloacetate
D-Aspartate
L-Asparaginyl-tRNA(Asn)
L-Aspartyl-tRNA(Asn)
tRNA(Asn)
2-Oxoglutarate
3-(4-Hydroxyphenyl)pyruvate
4-Hydroxy-2-oxoglutarate
Phenylpyruvate
Pyruvate
D-Glutamate
D-Phenylalanine
D-Erythrose
4-phosphate
2-Dehydro-3-deoxy-D-arabino-heptonate
7-phosphate
L-Homoserine
L-Threonine
O-Phospho-L-homoserine
3-Dehydroquinate
Methylglyoxal
RNA
sn-Glycerol
3-phosphate
Glycerone
phosphate
ATP
AMP
L-Homocysteine
5-Methyltetrahydropteroyltri-L-glutamate
L-Selenomethionine
L-Methionine
(S)-S-oxide
Tetrahydrofolate
Tetrahydropteroyltri-L-glutamate
L-Methionine
Selenohomocysteine
5-Methyltetrahydrofolate
S-Adenosyl-L-homocysteine
ADP
5'-Deoxyadenosine
dADP
Se-Adenosylselenomethionine
S-Adenosyl-L-methionine
Figure 3.7: Late Batch Pulse Experiment (
Gbatchpulse
(
s2
)), colour:green to red for
decreasing eccentricity, size: small to large for increasing stress
present. Membrane lipid synthesis/degradation and hydrofolate synthesis uniquely
belong to the H-graph.
Graph of the Solventogenesis
The solventogenic graph is larger than the acidogenic graph. Graph centralities are
shown in figure 3.10 which is again complemented with the
G
-graph, and
H
-graph
(figure 3.11). The most central metabolites are glutamate, ATP, further pyruvate
and phosphoenolpyruvate. In the medium and long range, a multitude of sugar
and nucleotide involving reactions are found. In contrast to acidogenic conditions,
carbon dioxide plays a central role together with pyruvate, ATP, ammonia and
glutamate. Surprisingly, also the position of butanoyl-CoA has changed, it is
shifted to the periphery. The production of carbondioxide is related to membrane
lipid conversion,, to pyruvate decarboxylation. Further reactions are
1.
rn:R06895: coproporphyrinogen-III:S-adenosyl-L-methionine oxidoreductase
3.3. DATA-DRIVEN PATHWAYS 49
5-Phospho-alpha-D-ribose
1-diphosphate
6,7-Dimethyl-8-(D-ribityl)lumazine
Xanthine
2,5-Diamino-6-(5-phospho-D-ribosylamino)pyrimidin-4(3H)-one
5-Amino-6-(1-D-ribitylamino)uracil
5-Amino-6-(5'-phospho-D-ribitylamino)uracil
5-Amino-6-(5'-phosphoribosylamino)uracil
Xanthosine
5'-phosphate
Riboflavin
2,3-Bisphospho-D-glycerate
ATP
beta-D-Glucose
AMP
beta-D-Glucose
6-phosphate
3-Phospho-D-glyceroyl
phosphate
ADP
dADP
Diphosphate
Deoxynucleoside
triphosphate
Adenosyl
cobinamide
dTTP
dGTP
dATP
dCTP
DNA
Adenosyl
cobinamide
phosphate
GDP-mannose
GDP
alpha-Ribazole
Cobamide
coenzyme
Dolichyl
phosphate
D-mannose
dGDP
D-Mannose
1-phosphate
GTP
Adenosine-GDP-cobinamide
Guanine
D-Ribose
5-phosphate
GMP
L-3,4-Dihydroxybutan-2-one
4-phosphate
D-Ribulose
5-phosphate
Formate
(S)-3-Hydroxybutanoyl-CoA
(R)-3-Hydroxybutanoyl-CoA
3-Hydroxybutanoyl-CoA
Ethanol
Propanoyl-CoA
Crotonoyl-CoA
Co(I) corrinoid
protein
2-Methylacetoacetyl-CoA
Acetaldehyde
CoA
CO
Acetyl-CoA
5-Methyl-5,6,7,8-tetrahydromethanopterin
3-Oxohexanoyl-CoA
Methyl-Co(III)
corrinoid protein
Acetoacetyl-CoA
5'-Deoxyadenosine
Butanoyl-CoA
Cobalt-precorrin 4
Butanal
S-Adenosyl-L-homocysteine
Cobalt-precorrin
5A
5,6,7,8-Tetrahydromethanopterin
sn-glycero-3-Phosphoethanolamine
Choline
tRNA(Sec)
sn-Glycerol
3-phosphate
Glycerol
Betaine aldehyde sn-glycero-3-Phosphocholine
Ethanolamine
Precorrin 4
DNA adenine
DNA
6-methylaminopurine
CMP
S-Adenosyl-L-methionine
Precorrin 5
L-Methionine
cis-Aconitate
Isocitrate
Citrate
CO2
2,3,4,5-Tetrahydrodipicolinate
Succinyl-CoA
2-Hydroxyglutarate
dTDP-4-dehydro-6-deoxy-L-mannose
dTDP-6-deoxy-L-mannose
Secondary alcohol
Ketone Aldehyde
1-Octanal
1-Octanol
Primary alcohol
3-Oxoacyl-[acyl-carrier
protein]
(3R)-3-Hydroxyacyl-[acyl-carrier
protein]
(3R)-3-Hydroxybutanoyl-[acyl-carrier
protein]
UDP-glucose
UDP-N-acetyl-D-galactosamine
UDP-N-acetyl-alpha-D-glucosamine
3-Oxodecanoyl-[acp]
2-Naphthaldehyde
1-Hydroxymethylnaphthalene
1-Naphthaldehyde
UDP-alpha-D-galactose
2-Methylprop-2-enoyl-CoA Phenylpropanoate
2-Methylpropanoyl-CoA
Acetoacetyl-[acp]
Ribonucleoside
diphosphate
NADH
2'-Deoxyribonucleoside
diphosphate
Retinal
2-(Formamido)-N1-(5'-phosphoribosyl)acetamidine
Retinol
Aminoimidazole
ribotide
trans-Cinnamate
NAD+
3-Ketopimeloyl-[acp]
methyl ester
3-Hydroxyglutaryl-[acp]
methyl ester
cis-3-Chloroallyl
aldehyde
2-Dehydro-3-deoxy-D-gluconate
D-Mannonate
cis-3-Chloro-2-propene-1-oltrans-3-Chloro-2-propene-1-ol
6-Deoxy-L-galactose
trans-3-Chloroallyl
aldehyde
L-Fucose
1-phosphate
3-Ketoglutaryl-[acp]
methyl ester
L-2,3-Dihydrodipicolinate
Pyruvate
Choline
phosphate
Inositol
1-phosphate
3-Hydroxypimeloyl-[acp]
methyl ester
Phosphatidylcholine
1-Phosphatidyl-D-myo-inositol
Phosphatidylethanolamine
3-Oxooctanoyl-[acp]
(3R)-3-Hydroxyoctanoyl-[acyl-carrier
protein]
(3S)-3-Hydroxyadipyl-CoA
Ethanolamine
phosphate
(2-Naphthyl)methanol
1-Alkenyl-2-acylglycerol
L-Aspartate
4-semialdehyde
(3R)-3-Hydroxydecanoyl-[acyl-carrier
protein]
3-Oxoadipyl-CoA
O-1-Alk-1-enyl-2-acyl-sn-glycero-3-phosphoethanolamine
Glycerophosphodiester
Phosphatidylglycerol
Alcohol
1,2-Diacyl-sn-glycerol
Cellobiose
3-Oxotetradecanoyl-[acp]
Dethiobiotin
(3R)-3-Hydroxypalmitoyl-[acyl-carrier
protein]
NADP+
3-Oxohexadecanoyl-[acp]
Cellulose
NADPH
dUDP
dCDP
(R)-3-Hydroxydodecanoyl-[acp]
Guanosine
3'-phosphate
CDP
2',3'-Cyclic CMP 3'-UMP
3-Oxododecanoyl-[acp]
2',3'-Cyclic GMP
3'-CMP
UDP
5-Phenyl-1,3-oxazinane-2,4-dione
4-Hydroxy-5-phenyltetrahydro-1,3-oxazin-2-one
Trypanothione
disulfide
2-Phenyl-1,3-propanediol
monocarbamate
3-Carbamoyl-2-phenylpropionaldehyde
Alcophosphamide
Trypanothione
Aldophosphamide
Trichloroethanol
Chloral hydrate
Dolichyl
phosphate
2,4-Diamino-6-nitrotoluene 3-Oxostearoyl-[acp]
3-Hydroxyoctadecanoyl-[acp]
2,4-Diamino-6-hydroxylaminotoluene
D-Fructose
1,6-bisphosphate
D-Erythrose
4-phosphate
Glycerone
phosphate
3-Phospho-D-glycerate
D-Glyceraldehyde
3-Phosphonooxypyruvate
D-Glyceraldehyde
3-phosphate
D-Fructose
beta-D-Fructose
1,6-bisphosphate
Sedoheptulose
1,7-bisphosphate
2-Phospho-D-glycerate
Hexanoyl-CoA
trans-Hex-2-enoyl-CoA
3alpha,7alpha,26-Trihydroxy-5beta-cholestane
3alpha,7alpha-Dihydroxy-5beta-cholestan-26-al (3R)-3-Hydroxytetradecanoyl-[acyl-carrier
protein]
Nucleoside
3'-phosphate
3,4-Dihydroxymandelaldehyde
3,4-Dihydroxyphenylethyleneglycol
2',3'-Cyclic UMP
(R)-3-Hydroxyhexanoyl-[acp]
2',3'-Cyclic AMP
3-Oxohexanoyl-[acp]
Thioredoxin
disulfide 2-Butenoate
3'-AMP
Thioredoxin
2',3'-Cyclic
nucleotide
Butanoic acid
2-Methylbut-2-enoyl-CoA
FADH2
dTDP-glucose
(S)-2-Methylbutanoyl-CoA dTDP-galactose
FAD Biotin
Maltose
6'-phosphate
UDP-N-acetylmuramoyl-L-alanyl-D-gamma-glutamyl-meso-2,6-diaminopimelate
L-Ascorbate
6-phosphate
alpha-D-Glucose
6-phosphate
Protein histidine
meso-2,6-Diaminoheptanedioate
Protein
N(pi)-phospho-L-histidine
alpha-D-Glucose
N-Acetyl-D-glucosamine
6-phosphate
N-Acetyl-D-galactosamine
6-phosphate
Galactitol
1-phosphate
Salicin
6-phosphate
D-Glucosamine
6-phosphate
Cobalt ion
N1-(5-Phospho-alpha-D-ribosyl)-5,6-dimethylbenzimidazole
Dimethylbenzimidazole
Nicotinate Nicotinate
D-ribonucleotide
Cobalt-sirohydrochlorin
Sirohydrochlorin UDP-N-acetylmuramate
UDP-N-acetylmuramoyl-L-alanine
L-Alanine
D-Galactosamine
6-phosphate Sorbitol
6-phosphate
Sorbose
1-phosphate
Sucrose
D-Mannitol
1-phosphate
L-Sorbose
D-Galactosamine
Sucrose
6-phosphate
10-Formyltetrahydrofolate
Tetrahydrofolate
5'-Phosphoribosyl-N-formylglycinamide
5'-Phosphoribosylglycinamide
5,10-Methenyltetrahydrofolate
Mannitol
Salicin
N-Acetyl-D-glucosamine
N-Acetyl-D-galactosamine
Ascorbate
Galactitol
D-Glucosamine
Lactose
6-phosphate
Lactose
alpha,alpha'-Trehalose
6-phosphate
Arbutin
6-phosphate
D-Sorbitol
N-Acetylmuramic
acid 6-phosphate
D-Mannose
6-phosphate
Maltose
Arbutin
D-Mannose
N-Acetylmuramate
alpha,alpha-Trehalose
D-Fructose
1-phosphate
Fumarate
N-(L-Arginino)succinate
Carbamoyl
phosphate
1-(5-Phospho-D-ribosyl)-5-amino-4-imidazolecarboxylate
Orthophosphate
L-Arginine
4-Phospho-L-aspartate
1-(5'-Phosphoribosyl)-5-amino-4-(N-succinocarboxamide)-imidazole
5-Carboxyamino-1-(5-phospho-D-ribosyl)imidazole
L-Aspartate
5'-Deoxy-5-fluorocytidine
5'-Deoxy-5-fluorouridine
5-Phosphoribosylamine
NH3
Uridine
Hydrogenobyrinate
Precorrin 8X
Cytidine
Deoxycytidine
Deoxyuridine
UTP
Cob(II)yrinate a,c
diamide Cobyrinate
c-monamide
CTP
L-Asparagine
Cobyrinate
L-Glutamine
Hydrogenobyrinate
a,c diamide
Cobalt-precorrin 8
LL-2,6-Diaminoheptanedioate
N-Acetyl-L-glutamate
N-Succinyl-2-L-amino-6-oxoheptanedioate
N-Succinyl-LL-2,6-diaminoheptanedioate
Succinate
L-Seryl-tRNA(Sec)
CDP-diacylglycerol
L-Seryl-tRNA(Ser)
Phosphatidylserine
UDP-N-acetylmuramoyl-L-alanyl-D-glutamate
L-Serine
D-Glucose
6-phosphate
L-Citrulline
D-Glucose
2-Phosphoglycolate
RNA
di-trans,poly-cis-Undecaprenyl
diphosphate
tRNA(Ser)
L-Ornithine
Glycolate
N-Acetyl-L-glutamate
5-phosphate
di-trans,poly-cis-Undecaprenyl
phosphate
3-Sulfinylpyruvate
3-Sulfopyruvate
L-Phenylalanine
L-Cysteine
3-(4-Hydroxyphenyl)pyruvate
L-Cysteate
3-Sulfino-L-alanine
L-Tyrosine
Mercaptopyruvate
D-4-Hydroxy-2-oxoglutarate
2-Oxoglutarate
DL-Glutamate
Phenylpyruvate
N-Acetyl-L-glutamate
5-semialdehyde
L-Glutamate
N-Acetylornithine
L-erythro-4-Hydroxyglutamate
Oxaloacetate
Figure 3.8: Continuous Experiment, Acidogenesis (
Hconti
(
s1
)), colour: green to
red for decreasing eccentricity, size: small to large for increasing stress
2.
rn:R03508: 1-(2-Carboxyphenylamino)-1-deoxy-D-ribulose-5-phosphate car-
boxy-lyase
3.
rn:R03348: Nicotinate-nucleotide:pyrophosphate phosphoribosyltransferase
4. rn:R01728: Prephenate:NAD+ oxidoreductase
5. rn:R01366: Acetoacetate carboxy-lyase
6. rn:R00965: orotidine-5’-phosphate carboxy-lyase
7. rn:R00451: meso-2,6-diaminoheptanedioate carboxy-lyase
8. rn:R00178: S-adenosyl-L-methionine carboxy-lyase
50 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
S-Adenosyl-L-homocysteine
Precorrin 5
CDP-diacylglycerol
S-Adenosyl-L-methionine
L-Methionine
5'-Deoxyadenosine
Phosphatidylserine
DNA
6-methylaminopurine
Alcohol
Glycerophosphodiester
sn-glycero-3-Phosphocholine
Ethanolamine
sn-glycero-3-Phosphoethanolamine
Glycerol
Phosphatidylglycerol
sn-Glycerol
3-phosphate
tRNA(Sec)
L-Seryl-tRNA(Ser)
tRNA(Ser)
L-Serine
Precorrin 4
L-Seryl-tRNA(Sec)
Salicin
6-phosphate
D-Glucosamine
6-phosphate
N-Acetyl-D-glucosamine
6-phosphate
Galactitol
1-phosphate
L-Sorbose
UDP-N-acetylmuramoyl-L-alanyl-D-glutamate
meso-2,6-Diaminoheptanedioate
UDP-N-acetylmuramoyl-L-alanyl-D-gamma-glutamyl-meso-2,6-diaminopimelate
N-Acetylmuramate
ADP
Maltose
beta-D-Glucose
6-phosphate
beta-D-Glucose
RNA
D-Fructose
2,3-Bisphospho-D-glycerate
3-Phospho-D-glyceroyl
phosphate
dADP
D-Glyceraldehyde
D-Fructose
1-phosphate
D-Mannose
1-phosphate
Dolichyl
phosphate
Adenosyl
cobinamide
dATP
dGDP
Dolichyl
phosphate
D-mannose
dTTP
GDP-mannose
DNA
D-Glyceraldehyde
3-phosphate
beta-D-Fructose
1,6-bisphosphate
D-Fructose
1,6-bisphosphate
Glycerone
phosphate
Sedoheptulose
1,7-bisphosphate
3-Phospho-D-glycerate
2-Phospho-D-glycerate
3-Phosphonooxypyruvate
D-Erythrose
4-phosphate
N-Acetylornithine
L-Aspartate
L-Ornithine
N-Acetyl-L-glutamate
5-semialdehyde
L-Citrulline
Acetyl-CoA
CoA
Succinyl-CoA
CO2
CO
CMPDNA adenine
D-4-Hydroxy-2-oxoglutarate
Succinate
Mercaptopyruvate
N-Succinyl-LL-2,6-diaminoheptanedioate
N-Acetyl-L-glutamate
LL-2,6-Diaminoheptanedioate
DL-Glutamate
2-Oxoglutarate
Phenylpyruvate
2,3,4,5-Tetrahydrodipicolinate
cis-Aconitate
Citrate
2-Hydroxyglutarate
N-Succinyl-2-L-amino-6-oxoheptanedioate
Isocitrate
L-Phenylalanine
3-Sulfinylpyruvate
3-Sulfopyruvate
L-Tyrosine
L-Cysteate
Oxaloacetate
L-Cysteine
L-Glutamate
3-Sulfino-L-alanine
L-erythro-4-Hydroxyglutamate
3-(4-Hydroxyphenyl)pyruvate
NH3
Cytidine
Deoxyuridine
Deoxycytidine
5'-Deoxy-5-fluorouridine
5'-Deoxy-5-fluorocytidine
Uridine
5-Phosphoribosylamine
1-(5'-Phosphoribosyl)-5-amino-4-(N-succinocarboxamide)-imidazole
Hydrogenobyrinate
a,c diamide
Cobyrinate
c-monamide
Cob(II)yrinate a,c
diamide
Fumarate
N-(L-Arginino)succinate
L-Glutamine CTP
UTP
L-Arginine
L-Asparagine
Cobyrinate
5-Carboxyamino-1-(5-phospho-D-ribosyl)imidazole
1-(5-Phospho-D-ribosyl)-5-amino-4-imidazolecarboxylate
Hydrogenobyrinate
Precorrin 8X
Cobalt-precorrin 8
Xanthine
5-Amino-6-(5'-phosphoribosylamino)uracil
Xanthosine
5'-phosphate
5-Phospho-alpha-D-ribose
1-diphosphate
2,5-Diamino-6-(5-phospho-D-ribosylamino)pyrimidin-4(3H)-one
Riboflavin
5-Amino-6-(1-D-ribitylamino)uracil
5-Amino-6-(5'-phospho-D-ribitylamino)uracil
Guanine
D-Ribose
5-phosphate
D-Ribulose
5-phosphate
L-3,4-Dihydroxybutan-2-one
4-phosphate
GMP
dCTP
AMP
dGTP
Cobamide
coenzyme
Deoxynucleoside
triphosphate
alpha-Ribazole
ATP
Diphosphate
Adenosyl
cobinamide
phosphate
GDP
Orthophosphate Formate
di-trans,poly-cis-Undecaprenyl
phosphate
2-Phosphoglycolate
di-trans,poly-cis-Undecaprenyl
diphosphate
GTP
Adenosine-GDP-cobinamide
Glycolate
4-Phospho-L-aspartate
N-Acetyl-L-glutamate
5-phosphate
Carbamoyl
phosphate
6,7-Dimethyl-8-(D-ribityl)lumazine
2'-Deoxyribonucleoside
diphosphate
trans-3-Chloroallyl
aldehyde
cis-3-Chloroallyl
aldehyde
Aminoimidazole
ribotide
Ribonucleoside
diphosphate
2-Dehydro-3-deoxy-D-gluconatecis-3-Chloro-2-propene-1-ol
D-Mannonate
1-Octanol
(3R)-3-Hydroxyacyl-[acyl-carrier
protein]
Phenylpropanoate
3-Oxoacyl-[acyl-carrier
protein]
1-Octanal
Retinol
trans-Cinnamate2-Methylpropanoyl-CoA
dTDP-4-dehydro-6-deoxy-L-mannose
dTDP-6-deoxy-L-mannose
2-Methylprop-2-enoyl-CoA
3-Ketoglutaryl-[acp]
methyl ester
3-Hydroxyglutaryl-[acp]
methyl ester
3-Ketopimeloyl-[acp]
methyl ester
Pyruvate
L-Aspartate
4-semialdehyde
L-2,3-Dihydrodipicolinate
3-Hydroxypimeloyl-[acp]
methyl ester
2-Naphthaldehyde
3-Oxoadipyl-CoA
1-Naphthaldehyde
(2-Naphthyl)methanol 1-Hydroxymethylnaphthalene
(3S)-3-Hydroxyadipyl-CoA
UDP-N-acetyl-D-galactosamine
UDP-N-acetyl-alpha-D-glucosamine
UDP-alpha-D-galactosePrimary alcohol
AldehydeKetone
NADH
UDP-glucose
NAD+
Secondary alcohol
Acetoacetyl-[acp]
3-Oxodecanoyl-[acp]3-Oxooctanoyl-[acp]
(3R)-3-Hydroxydecanoyl-[acyl-carrier
protein]
(3R)-3-Hydroxybutanoyl-[acyl-carrier
protein]
(3R)-3-Hydroxyoctanoyl-[acyl-carrier
protein]
L-Alanine
5,10-Methenyltetrahydrofolate
10-Formyltetrahydrofolate
Tetrahydrofolate
5'-Phosphoribosyl-N-formylglycinamide
5'-Phosphoribosylglycinamide
Cobalt-sirohydrochlorin UDP-N-acetylmuramoyl-L-alanine
Cobalt ion
UDP-N-acetylmuramateSirohydrochlorin
D-Galactosamine D-Sorbitol
Dimethylbenzimidazole
Nicotinate
D-ribonucleotide
Sucrose
Nicotinate
N1-(5-Phospho-alpha-D-ribosyl)-5,6-dimethylbenzimidazole
Arbutin
6-phosphate
Lactose
alpha,alpha'-Trehalose
6-phosphate
Sorbitol
6-phosphate
D-Mannose
Arbutin
alpha,alpha-Trehalose
D-Mannose
6-phosphate
Lactose
6-phosphate
D-Mannitol
1-phosphateSorbose
1-phosphate
N-Acetyl-D-galactosamine
6-phosphate
Sucrose
6-phosphate
L-Ascorbate
6-phosphate
Protein
N(pi)-phospho-L-histidine
Protein histidine
D-Galactosamine
6-phosphate
Ascorbate
N-Acetyl-D-glucosamine
Galactitol
D-Glucosamine
N-Acetyl-D-galactosamine
Salicin
Mannitol
L-Fucose
1-phosphate
2-(Formamido)-N1-(5'-phosphoribosyl)acetamidine
Retinal
6-Deoxy-L-galactose
trans-3-Chloro-2-propene-1-ol
alpha-D-Glucose
6-phosphate
D-Glucose
alpha-D-Glucose
N-Acetylmuramic
acid 6-phosphate
Maltose
6'-phosphate
D-Glucose
6-phosphate
5,6,7,8-Tetrahydromethanopterin
Acetoacetyl-CoA
Butanal
5-Methyl-5,6,7,8-tetrahydromethanopterin
Butanoyl-CoA
3-Oxohexanoyl-CoA
Cobalt-precorrin
5A
Cobalt-precorrin 4
Methyl-Co(III)
corrinoid protein
Propanoyl-CoA
Ethanol
2-Methylacetoacetyl-CoA
Co(I) corrinoid
protein
Acetaldehyde
3-Hydroxybutanoyl-CoA
Crotonoyl-CoA
(S)-3-Hydroxybutanoyl-CoA
(R)-3-Hydroxybutanoyl-CoA
Phosphatidylethanolamine
Phosphatidylcholine
Choline
phosphate
1-Phosphatidyl-D-myo-inositol
1,2-Diacyl-sn-glycerol
Betaine aldehyde Choline
O-1-Alk-1-enyl-2-acyl-sn-glycero-3-phosphoethanolamine
1-Alkenyl-2-acylglycerol
Inositol
1-phosphate
Ethanolamine
phosphate
3'-CMP
dUDP
3'-UMP
Guanosine
3'-phosphate
2',3'-Cyclic CMP
(R)-3-Hydroxydodecanoyl-[acp]
3-Oxododecanoyl-[acp]
CDP UDP
dCDP
2',3'-Cyclic GMP
2-Methylbut-2-enoyl-CoA
Nucleoside
3'-phosphate
dTDP-glucose
2',3'-Cyclic
nucleotide
Cellobiose
CellulosedTDP-galactose(S)-2-Methylbutanoyl-CoA
3'-AMP
(R)-3-Hydroxyhexanoyl-[acp]
2',3'-Cyclic AMP
3-Oxohexanoyl-[acp]
3,4-Dihydroxyphenylethyleneglycol
3,4-Dihydroxymandelaldehyde
2',3'-Cyclic UMP
3-Oxotetradecanoyl-[acp] 3-Oxohexadecanoyl-[acp]
(3R)-3-Hydroxypalmitoyl-[acyl-carrier
protein]
3alpha,7alpha,26-Trihydroxy-5beta-cholestaneHexanoyl-CoA
(3R)-3-Hydroxytetradecanoyl-[acyl-carrier
protein]
3alpha,7alpha-Dihydroxy-5beta-cholestan-26-altrans-Hex-2-enoyl-CoA
Thioredoxin
disulfide
Thioredoxin NADP+
NADPH
Butanoic acid FAD Dethiobiotin
Biotin
FADH22-Butenoate
3-Oxostearoyl-[acp]Aldophosphamide
Trichloroethanol
Chloral hydrate3-Carbamoyl-2-phenylpropionaldehyde 2,4-Diamino-6-nitrotoluene5-Phenyl-1,3-oxazinane-2,4-dione
Trypanothione
disulfide
3-Hydroxyoctadecanoyl-[acp]
2,4-Diamino-6-hydroxylaminotoluene
Alcophosphamide
2-Phenyl-1,3-propanediol
monocarbamate
4-Hydroxy-5-phenyltetrahydro-1,3-oxazin-2-oneTrypanothione
Figure 3.9: Continuous Experiment, Acidogenesis, with Augmentation, green
and red: Hconti(s1), red: Gconti(s1)
One finally observes several sulfuric aminoacid relevant reactions in the H-graph.
3.3. DATA-DRIVEN PATHWAYS 51
Acetone
cyanohydrin
(2R)-2-Hydroxy-2-methylbutanenitrile
(S)-4-Hydroxymandelonitrile
Alcohol
Fe2+
D-Glucose
6-phosphate
beta-D-Glucose
Lactose
6-phosphate
Ascorbate
D-Galactose
6-phosphate
beta-D-Fructose
6-Phospho-beta-D-glucosyl-(1,4)-D-glucose
Galactitol
Cellobiose
2-Naphthaldehyde
2,5-Diamino-6-(5'-triphosphoryl-3',4'-trihydroxy-2'-oxopentyl)-amino-4-oxopyrimidine
2,5-Diaminopyrimidine
nucleoside
triphosphate
6-Pyruvoyltetrahydropterin
Triphosphate
D-Mannose
1-phosphate 1-Naphthaldehyde
(2-Naphthyl)methanol
3-Keto-beta-D-galactose
Linamarin
6-Phospho-beta-D-galactoside
Lotaustralin
Dhurrin 3-Ketolactose
Cellulose
Precorrin 5
Phospholipid
cyclopropane
fatty acid
Precorrin 2
S-Adenosyl-L-homocysteine
Cobalt-factor III
tRNA(Met)
S-Adenosylmethioninamine
5'-Deoxyadenosine
Precorrin 3A
Sirohydrochlorin
Cobalt-precorrin
5A
D-Glucose
D-Glucoside
cis-beta-D-Glucosyl-2-hydroxycinnamate Cyanohydrin
Cyanoglycoside
cis-2-Hydroxycinnamate
Prunasin
Amygdalin
Mandelonitrile
ROH
Phospholipid
olefinic fatty acid
Siroheme
L-Selenomethionine
Cobalt-sirohydrochlorin
Precorrin 4
Selenomethionyl-tRNA(Met)
Uroporphyrin III
Cobalt-precorrin 4
Arbutin
Salicyl alcohol
beta-D-Glucose
6-phosphate
Hydroquinone
Salicin
N-Acetyl-D-galactosamine
6-phosphate
Salicin
6-phosphate
D-Galactosyl-N-acetyl-D-galactosaminyl-(N-acetylneuraminyl)-D-galactosyl-D-glucosylceramide
D-Mannose
D-Galactosamine
GA1
Arbutin
6-phosphate
beta-D-Galactosyl-(1->4)-beta-D-glucosyl-(1<->1)-ceramide
N-Acetyl-D-glucosamine
Sorbose
1-phosphate
D-Sorbitol
alpha-D-Glucose
6-phosphate
alpha,alpha-Trehalose
Lactose
Sorbitol
6-phosphate
L-Sorbose
D-Glucosamine
6-phosphate
D-Glucosamine
D-Fructose
Maltose
6'-phosphate
L-Ascorbate
6-phosphate
Sucrose
alpha-D-Glucose
Maltose
Galactitol
1-phosphate
Sucrose
6-phosphate
N-Acetylmuramic
acid 6-phosphate
N-Acetyl-D-glucosamine
6-phosphate
D-Galactose
N-Acetylmuramate
Galactan
GA2
N-Acetyl-D-galactosaminyl-(N-acetylneuraminyl)-D-galactosyl-D-glucosylceramide
alpha-D-Galactosyl-(1->4)-beta-D-galactosyl-(1->4)-beta-D-glucosyl-(1<->1)-ceramide
Globoside
GM3
beta-D-Glucosyl-(1<->1)-ceramide
Chitobiose
N-Acetyl-D-galactosamine
alpha,alpha'-Trehalose
6-phosphate
Protein
N(pi)-phospho-L-histidine
D-Mannose
6-phosphate
D-Galactosamine
6-phosphate
Protein histidine
alpha-D-Galactose
1-phosphate
D-Glyceraldehyde
3-phosphate
Orotidine
5'-phosphate
beta-D-Fructose
6-phosphate
3-Phospho-D-glycerate
Hydroxymethylbilane
2-Dehydro-3-deoxy-D-arabino-heptonate
7-phosphate
beta-D-Fructose
1,6-bisphosphate
D-Fructose
6-phosphate
Phosphoenolpyruvate
Glycerone
phosphate
Mannitol
D-Erythrose
4-phosphate
Precorrin 1
beta-D-Glucose
1-phosphate
L-Methionine
S-Adenosyl-L-methionine
Uroporphyrinogen
III
Methanethiol
UDP-alpha-D-galactose
D-Mannitol
1-phosphate
D-Fructose
1-phosphate
UDP-glucose
Sedoheptulose
7-phosphate
3-Dehydroquinate
L-Methionyl-tRNA
Indoleglycerol
phosphate
Orthophosphate
Indole
N-Acetyl-L-glutamate
5-phosphate
Shikimate
3-phosphate
D-Glucose
1-phosphate
Shikimate
3-Dehydroshikimate
Folinic acid
Sedoheptulose
1,7-bisphosphate
N-Acetyl-L-glutamate
5-semialdehyde
ATP
D-Fructose
1,6-bisphosphate
L-Tryptophanyl-tRNA(Trp)
ADP-glucose
L-Tryptophan
Uridine
Starch
1-Octanal
1-Octanol
tRNA(Trp)
dTDP-glucose
dTDP-galactose
5,10-Methenyltetrahydrofolate
1-Hydroxymethylnaphthalene
L-Phenylalanine
D-4-Hydroxy-2-oxoglutarate
3-Sulfino-L-alanine
L-Cysteate
L-Isoleucine
L-Homoserine
O-Succinyl-L-homoserine
N-Acetylornithine
Pyruvate
[Enzyme]-S-sulfanylcysteine
L-Cysteine
L-Glutamate
L-Valine
DL-Glutamate
2-Oxoglutarate
Mercaptopyruvate
3-Methyl-2-oxobutanoic
acid
D-Phenylalanine
L-Leucyl-tRNA
4-Methyl-2-oxopentanoate
L-Isoleucyl-tRNA(Ile)
L-Leucine
3-Sulfopyruvate
Phenylpyruvate
(R)-2,3-Dihydroxy-3-methylpentanoate
3-Sulfinylpyruvate
tRNA(Ile)
(S)-3-Methyl-2-oxopentanoic
acid
(R)-2,3-Dihydroxy-3-methylbutanoate
tRNA(Leu)
L-Lysine
meso-2,6-Diaminoheptanedioate
2,3-Dihydroxy-3-methylbutanoate
Coproporphyrinogen
III
Protoporphyrinogen
IX
D-Glyceraldehyde
Dihydrofolate
D-Glycerate
L-Fucose
1-phosphate
2'-Deoxyribonucleoside
triphosphate
D-erythro-1-(Imidazol-4-yl)glycerol
3-phosphate
Hydrogen
peroxide
Ribonucleoside
triphosphate
2'-Deoxyribonucleoside
diphosphate
Oxygen
Ribonucleoside
diphosphate
3-(Imidazol-4-yl)-2-oxopropyl
phosphate
15(S)-HPETE
tRNA(Pro)
L-Proline
L-Prolyl-tRNA(Pro)
L-Threonine
5(S)-HETE
L-Threonyl-tRNA(Thr)
tRNA(Thr)
(15S)-15-Hydroxy-5,8,11-cis-13-trans-eicosatetraenoate
5(S)-HPETE
Xanthine
Tetrahydrofolate
Hypoxanthine
5'-Deoxy-5-fluorouridine
Adenosine-GDP-cobinamide
Guanine
Adenine
5'-Deoxy-5-fluorocytidine
Diimine
Nitrogen
Hydrazine
Deoxycytidine
Deoxyuridine
Glutathione
disulfide
tRNA(Asp)
L-Arginine
tRNA(Arg)
L-Histidinal
L-Histidine
L-Arginyl-tRNA(Arg)
NAD+
NADH
L-Histidinol
Methaneselenol
Tetrahydrofolyl-[Glu](2)
Glutathione
S-Glutathionyl-L-cysteine
L-Aspartate
tRNA(Asn)
L-Aspartyl-tRNA(Asn)
Fumarate
L-Alanine
L-Aspartyl-tRNA(Asp)
Dihydropteroate
D-Aspartate
10-Formyltetrahydrofolylpolyglutamate
Tetrahydrofolyl-[Glu](n)
L-erythro-4-Hydroxyglutamate
10-Formyltetrahydrofolyl
L-glutamate
10-Formyltetrahydrofolate
6-Deoxy-L-galactose
Oxaloacetate
L-Ornithine
Methylselenic acid
Cobalt-precorrin 6 Ethylene
Acetylene
Arsenite cis-3-Chloroallyl
aldehyde
Arsenate ion cis-3-Chloro-2-propene-1-ol
Cobalt-dihydro-precorrin
6
Retinol
Orotate
Retinal
Thioredoxin
Thioredoxin
disulfide
7,8-Dihydroneopterin
3'-triphosphate
GDP
Formamidopyrimidine
nucleoside
triphosphate
Formate
(S)-Dihydroorotate
GDP-mannose
Thiamin
monophosphate
dTDP
N-((R)-Pantothenoyl)-L-cysteine
4-Methyl-5-(2-phosphoethyl)-thiazole
5-Fluorouridine
Pantetheine
dCDP
dGDP
CDP
dUDP
UDP
D-Xylulose
IDP GTP
(S)-2-Aceto-2-hydroxybutanoate
2-Oxobutanoate
(R)-Lactate
2-Acetolactate
(S)-2-Acetolactate
Succinyl-CoA
Prephenate
2-Deoxy-D-ribose
5-phosphate
CO2
Ethanol
2-(alpha-Hydroxyethyl)thiamine
diphosphate
Acetaldehyde
D-Alanine
Indolepyruvate
L-Aspartate
4-semialdehyde
Thiamin
diphosphate
L-2,3-Dihydrodipicolinate
2-Oxo acid
Acetone
CoA
5-Guanidino-2-oxopentanoate
S-2-(Indol-3-yl)acetyl-CoA
Acetoacetyl-CoA
(S)-2-Hydroxyacid
D-Amino acid
Acetoacetate
alpha-Isopropylmalate
Acetyl-CoA
Aldehyde
2-Isopropylmaleate
3-Oxohexanoyl-CoA
5-Amino-2-oxopentanoic
acid
Butanoyl-CoA
D-Lysine
D-Arginine
D-Ornithine
6-Amino-2-oxohexanoate
(2R,3S)-3-Isopropylmalate
2-Methylacetoacetyl-CoA
Ribosomal-protein
N-acetyl-L-alanine
Primary alcohol
Propanoyl-CoA
Butanoic acid
Butanal
Ribosomal-protein
L-alanine
3-Oxohexanoyl-[acp]
Malonyl-[acyl-carrier
protein]
3-Ketoglutaryl-[acp]
methyl ester
3-Oxodecanoyl-[acp]
3-Oxotetradecanoyl-[acp]
Acetoacetyl-[acp]
3-Oxododecanoyl-[acp]
3-Oxooctanoyl-[acp]
3-Oxohexadecanoyl-[acp]
3-Ketopimeloyl-[acp]
methyl ester
3-Oxostearoyl-[acp]
Hexanoyl-[acp]
Acetyl-[acyl-carrier
protein]
Dodecanoyl-[acyl-carrier
protein]
Decanoyl-[acp]
Tetradecanoyl-[acp]
Malonyl-[acp]
methyl ester
Octanoyl-[acp]
Butyryl-[acp]
Acyl-carrier
protein
Glutaryl-[acp]
methyl ester
Hexadecanoyl-[acp]
L-Tyrosine
N-Acetyl-L-glutamate
3-(4-Hydroxyphenyl)pyruvate
D-Glutamate
[Enzyme]-cysteine
trans-3-Chloro-2-propene-1-ol
Ferricytochrome c
Choline
NADPH
5-Aminolevulinate
D-Gluconic acid
Trypanothione
5-Dehydro-D-gluconate
(S)-4-Amino-5-oxopentanoate
NADP+
Ferrocytochrome
c
Trypanothione
disulfide
Ketone
D-Glucuronate
4-Hydroxy-5-phenyltetrahydro-1,3-oxazin-2-one
3,4-Dihydroxyphenylethyleneglycol
3,4-Dihydroxymandelaldehyde
L-Glutamyl-tRNA(Glu)
D-Galacturonate
D-Tagaturonate
Molybdoenzyme
molybdenum
cofactor
Molybdate
Precorrin 6Y
trans-3-Chloroallyl
aldehyde
FADH2
Precorrin 6X
Betaine aldehyde
N-(5'-Phospho-D-1'-ribulosylformimino)-5-amino-1-(5''-phospho-D-ribosyl)-4-imidazolecarboxamide
Secondary alcohol
3alpha,7alpha-Dihydroxy-5beta-cholestan-26-al
Glycolate
UDP-N-acetyl-alpha-D-glucosamine
Glyoxylate
UDP-N-acetyl-D-galactosamine
3alpha,7alpha,26-Trihydroxy-5beta-cholestane5-(5-Phospho-D-ribosylaminoformimino)-1-(5-phosphoribosyl)-imidazole-4-carboxamide
GDP-L-fucose
Aldophosphamide
D-Fructuronate
3-Carbamoyl-2-phenylpropionaldehyde
5-Phenyl-1,3-oxazinane-2,4-dione
GDP-4-dehydro-6-deoxy-D-mannose
2-Phenyl-1,3-propanediol
monocarbamate
Alcophosphamide
GDP-4-keto-6-L-deoxygalactose
tRNA(Glu)
(R)-2-Methylmalate
Trichloroethanol
2-Methylmaleate
D-erythro-3-Methylmalate
Chloral hydrate
Xanthosine
5'-phosphate
IMP
Adenylated
molybdopterin
Cobamide
coenzyme
alpha-Ribazole
6-Thioinosine-5'-monophosphate
L-Cystine
L-Cystathionine
Thiocysteine
Selenohomocysteine
Chorismate
Succinate
Hydrogen sulfide
L-Homocysteine
Cystathionine
L-Glutamine
N-Carbamoyl-L-aspartate
NH3
Thiosulfate
Anthranilate
S-Sulfo-L-cysteine
L-Citrulline
Hydrogen
selenide
Nitrite
HCO3-
GMP
Sulfite
6-Thioguanosine
monophosphate
O-Phosphorylhomoserine
Quinolinate
Cytidine Nicotinate
D-ribonucleotide
5-O-(1-Carboxyvinyl)-3-phosphoshikimate
1-(2-Carboxyphenylamino)-1-deoxy-D-ribulose
5-phosphate
N-(5-Phospho-D-ribosyl)anthranilate
Porphobilinogen
1-(5-Phospho-D-ribosyl)-ATP 5-Phospho-alpha-D-ribose
1-diphosphate
O-Acetyl-L-homoserine
L-Selenocystathionine
O-Acetyl-L-serine
Iminoaspartate
L-Selenocysteine
Acetate
Carbamoyl
phosphate
ADP
CMP
UTP
CTP
D-Xylulose
5-phosphate
dGTP ITP
UMP
dATP
Diphosphate
DNA D-Tagatose
1,6-bisphosphate
dCTP
dTTP
dUTP
Deoxynucleoside
triphosphate
(R)-4'-Phosphopantothenoyl-L-cysteine
2-Dehydro-3-deoxy-6-phospho-D-gluconate
Pantothenate
Pantetheine
4'-phosphate
dADP
2-Methyl-4-amino-5-hydroxymethylpyrimidine
diphosphate
D-4'-Phosphopantothenate
D-Tagatose
6-phosphate
5-Fluorouridine
monophosphate
Glycine
5,10-Methylenetetrahydromethanopterin
2-Dehydro-3-deoxy-D-gluconate
D-Altronate
Selenate
5,6,7,8-Tetrahydromethanopterin
FAD
Adenylyl sulfate
3'-Phosphoadenylylselenate
Sulfate Selenite
5,10-Methylenetetrahydrofolate
Adenylylselenate
6-Thioxanthine
5'-monophosphate
3'-Phosphoadenylyl
sulfate
L-Serine
AMP
Figure 3.10: Continous Experiment, Solventogenesis (
Hconti
(
s1
)), colour:green to
red for decreasing eccentricity, size: small to large for increasing stress
52 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
(S)-4-Hydroxymandelonitrile 3-Keto-beta-D-galactose
(2R)-2-Hydroxy-2-methylbutanenitrile
Fe2+
D-Galactose
6-phosphate
beta-D-Fructose
Ascorbate Galactitol
Lactose
6-phosphate
6-Phospho-beta-D-glucosyl-(1,4)-D-glucose
beta-D-Glucose
Cellobiose
D-Glucose
6-phosphate
2,5-Diaminopyrimidine
nucleoside
triphosphate
6-Pyruvoyltetrahydropterin
2,5-Diamino-6-(5'-triphosphoryl-3',4'-trihydroxy-2'-oxopentyl)-amino-4-oxopyrimidine
Triphosphate
D-Mannose
1-phosphate 2-Naphthaldehyde
(2-Naphthyl)methanol 1-Naphthaldehyde
3-Ketolactose
Dhurrin
Acetone
cyanohydrin
Linamarin
Lotaustralin
Alcohol
6-Phospho-beta-D-galactoside Cellulose
Cobalt-precorrin
5A
Sirohydrochlorin
Phospholipid
cyclopropane
fatty acid
Precorrin 5
S-Adenosylmethioninamine
Precorrin 3A
5'-Deoxyadenosine
S-Adenosyl-L-homocysteine
Precorrin 2
tRNA(Met)
Cobalt-factor III
ROH
D-Glucoside
Cyanohydrin
cis-2-Hydroxycinnamate
cis-beta-D-Glucosyl-2-hydroxycinnamate
Mandelonitrile
Amygdalin
Cyanoglycoside
Prunasin
D-Glucose
Selenomethionyl-tRNA(Met)
L-Selenomethionine
Precorrin 4
Uroporphyrin III
Cobalt-precorrin 4
Siroheme
Phospholipid
olefinic fatty acid
Cobalt-sirohydrochlorin
Salicin
Hydroquinone
Arbutin
Salicyl alcohol
beta-D-Glucose
6-phosphate
D-Galactosamine
Arbutin
6-phosphate
Salicin
6-phosphate
alpha,alpha-Trehalose
D-Mannose
D-Galactosyl-N-acetyl-D-galactosaminyl-(N-acetylneuraminyl)-D-galactosyl-D-glucosylceramide
N-Acetyl-D-glucosamine
GA1
beta-D-Galactosyl-(1->4)-beta-D-glucosyl-(1<->1)-ceramide
alpha-D-Glucose
alpha-D-Glucose
6-phosphate
L-Sorbose
D-Galactose
Protein
N(pi)-phospho-L-histidine
D-Sorbitol
Protein histidine
alpha,alpha'-Trehalose
6-phosphate
Lactose
N-Acetyl-D-galactosamine
6-phosphate
D-Mannose
6-phosphate
D-Galactosamine
6-phosphate
Sorbitol
6-phosphate
Sorbose
1-phosphate
Protoporphyrinogen
IX
N-Acetylmuramate
N-Acetyl-D-glucosamine
6-phosphate
Galactan
N-Acetylmuramic
acid 6-phosphate
alpha-D-Galactosyl-(1->4)-beta-D-galactosyl-(1->4)-beta-D-glucosyl-(1<->1)-ceramide
GM3
Globoside
N-Acetyl-D-galactosaminyl-(N-acetylneuraminyl)-D-galactosyl-D-glucosylceramide
Chitobiose
GA2
beta-D-Glucosyl-(1<->1)-ceramide
N-Acetyl-D-galactosamine
D-Glucosamine
D-Fructose Maltose
Galactitol
1-phosphate
D-Glucosamine
6-phosphate
Sucrose
6-phosphate
Maltose
6'-phosphate
D-Fructose
1-phosphate
Sucrose
L-Ascorbate
6-phosphate
Formate
Formamidopyrimidine
nucleoside
triphosphate
GDP-mannose
D-Xylulose
3'-Phosphoadenylylselenate
ITP
ADP
CMP
Diphosphate
D-Xylulose
5-phosphate
3'-Phosphoadenylyl
sulfate
Adenylylselenate
Adenylyl sulfate
AMP
L-Serine
Sulfate
Trichloroethanol
2-Methylmaleate
(R)-2-Methylmalate
2'-Deoxyribonucleoside
triphosphate
Chloral hydrate
D-erythro-3-Methylmalate
dGTP
D-Tagatose
1,6-bisphosphate
CTP
dATP
D-4-Hydroxy-2-oxoglutarate
L-Phenylalanine
L-Cysteate
3-Sulfopyruvate
3-Sulfinylpyruvate
10-Formyltetrahydrofolylpolyglutamate
L-Leucine
Phenylpyruvate
L-Leucyl-tRNA
L-Isoleucine
(S)-3-Methyl-2-oxopentanoic
acid
4-Methyl-2-oxopentanoate
(R)-2,3-Dihydroxy-3-methylpentanoate
tRNA(Ile)
L-Isoleucyl-tRNA(Ile)
D-Glyceraldehyde
L-Aspartate
L-Alanine
Fumarate
Acetate
L-Ornithine
L-Lysine
Acetoacetate
Ethanol
2-Oxo acid
S-2-(Indol-3-yl)acetyl-CoA
(R)-2,3-Dihydroxy-3-methylbutanoate
2,3-Dihydroxy-3-methylbutanoate
tRNA(Leu)
D-Aspartate
10-Formyltetrahydrofolate
Methaneselenol
Methylselenic acid
L-erythro-4-Hydroxyglutamate
Tetrahydrofolyl-[Glu](n)
Dihydropteroate
L-Fucose
1-phosphate
Dihydrofolate
beta-D-Glucose
1-phosphate
3-Phospho-D-glycerate
Phosphoenolpyruvate
alpha-D-Galactose
1-phosphate
UDP-alpha-D-galactose
Orotidine
5'-phosphate
D-Fructose
6-phosphate
2-Dehydro-3-deoxy-D-arabino-heptonate
7-phosphate
N-Acetylornithine
2-Oxoglutarate
D-Glutamate
D-Phenylalanine
[Enzyme]-S-sulfanylcysteine
3-Methyl-2-oxobutanoic
acid
L-Glutamate
L-Valine
DL-Glutamate
Selenohomocysteine
L-Homocysteine
Quinolinate
O-Acetyl-L-homoserine
Chorismate
L-Selenocystathionine
Cystathionine
Hydrogen sulfide
L-Homoserine
L-Tyrosine
Mercaptopyruvate
3-Sulfino-L-alanine
L-Cystathionine
Succinate
Oxaloacetate
10-Formyltetrahydrofolyl
L-glutamate
6-Deoxy-L-galactose
Ribonucleoside
triphosphate
Oxygen
2'-Deoxyribonucleoside
diphosphate
3-(Imidazol-4-yl)-2-oxopropyl
phosphate
D-erythro-1-(Imidazol-4-yl)glycerol
3-phosphate
Hydrogen
peroxide
Ribonucleoside
diphosphate
5(S)-HPETE
L-Threonyl-tRNA(Thr)
L-Threonine
tRNA(Pro)
5(S)-HETE
15(S)-HPETE
tRNA(Thr)
L-Proline
(15S)-15-Hydroxy-5,8,11-cis-13-trans-eicosatetraenoate
L-Prolyl-tRNA(Pro)
Adenosine-GDP-cobinamide
Hypoxanthine
5'-Deoxy-5-fluorouridine
Xanthine
Tetrahydrofolate
Hydrazine
Adenine
Guanine
5'-Deoxy-5-fluorocytidine
Deoxycytidine
Diimine
Deoxyuridine
L-Arginine
tRNA(Arg)
Glutathione
disulfide
tRNA(Asp)
tRNA(Asn)
NADH
NAD+
L-Histidine
L-Histidinal
L-Histidinol
L-Arginyl-tRNA(Arg)
D-Glycerate
S-Glutathionyl-L-cysteine
Tetrahydrofolyl-[Glu](2)
L-Aspartyl-tRNA(Asn)
L-Aspartyl-tRNA(Asp)
Glutathione
Cobalt-dihydro-precorrin
6
Arsenate ion
Ethylene
cis-3-Chloro-2-propene-1-ol
Cobalt-precorrin 6 Arsenite cis-3-Chloroallyl
aldehyde
Acetylene
Retinol (S)-Dihydroorotate
Thioredoxin
Thioredoxin
disulfide
Retinal
Orotate
UTP
GTP
IDP
GDP
7,8-Dihydroneopterin
3'-triphosphate
Pantothenate
2-Dehydro-3-deoxy-6-phospho-D-gluconate
(R)-4'-Phosphopantothenoyl-L-cysteine
D-4'-Phosphopantothenate
[Enzyme]-cysteine
2-Oxobutanoate
Pyruvate
2-Deoxy-D-ribose
5-phosphate
(S)-2-Aceto-2-hydroxybutanoate
2-Acetolactate
(S)-2-Acetolactate
CO2
N-Acetyl-L-glutamate
3-(4-Hydroxyphenyl)pyruvate
Prephenate
meso-2,6-Diaminoheptanedioate
Coproporphyrinogen
III
L-Aspartate
4-semialdehyde
D-Alanine
L-2,3-Dihydrodipicolinate
(R)-Lactate
Acetaldehyde
Indolepyruvate
2-(alpha-Hydroxyethyl)thiamine
diphosphate
Thiamin
diphosphate
Succinyl-CoA
alpha-Isopropylmalate
D-Amino acid
Acetyl-CoA (S)-2-Hydroxyacid
CoA
5-Guanidino-2-oxopentanoate
Acetone
Acetoacetyl-CoA
2-Isopropylmaleate
(2R,3S)-3-Isopropylmalate
D-Arginine
D-Ornithine
Aldehyde
5-Amino-2-oxopentanoic
acid
3-Oxohexanoyl-CoA
Butanoyl-CoA
D-Lysine
6-Amino-2-oxohexanoate
Butanal
Ribosomal-protein
L-alanine
Ribosomal-protein
N-acetyl-L-alanine
Primary alcohol
Propanoyl-CoA
Butanoic acid
2-Methylacetoacetyl-CoA
3-Oxooctanoyl-[acp]
Acetoacetyl-[acp]
3-Oxostearoyl-[acp]
3-Oxododecanoyl-[acp]
3-Oxohexanoyl-[acp]
3-Oxotetradecanoyl-[acp]
3-Oxohexadecanoyl-[acp]
3-Oxodecanoyl-[acp]
Malonyl-[acyl-carrier
protein]
3-Ketoglutaryl-[acp]
methyl ester
3-Ketopimeloyl-[acp]
methyl ester
Octanoyl-[acp]
Malonyl-[acp]
methyl ester
Butyryl-[acp]
Dodecanoyl-[acyl-carrier
protein]
Acyl-carrier
protein
Glutaryl-[acp]
methyl ester
Decanoyl-[acp]
Hexanoyl-[acp]
Acetyl-[acyl-carrier
protein]
Hexadecanoyl-[acp]
Tetradecanoyl-[acp]
NADPH
Ferricytochrome c
5-Aminolevulinate
Trypanothione
trans-3-Chloro-2-propene-1-ol
Choline
D-Gluconic acid
D-Tagaturonate
5,6,7,8-Tetrahydromethanopterin
Thiamin
monophosphate
D-Galacturonate
4-Methyl-5-(2-phosphoethyl)-thiazole
NADP+
Ferrocytochrome
c
Trypanothione
disulfide
(S)-4-Amino-5-oxopentanoate
5-Dehydro-D-gluconate
D-Altronate
5,10-Methylenetetrahydromethanopterin
Glycine
2-Dehydro-3-deoxy-D-gluconate
N-((R)-Pantothenoyl)-L-cysteine
Selenate
Pantetheine
dTDP
dADP
dUDP
5-Fluorouridine
monophosphate
Pantetheine
4'-phosphate
5-Fluorouridine
2-Methyl-4-amino-5-hydroxymethylpyrimidine
diphosphate
dGDP
UDP
dUTP
DNA
dTTP
dCDP
Deoxynucleoside
triphosphate
D-Tagatose
6-phosphate dCTP
CDP
GDP-4-keto-6-L-deoxygalactose
tRNA(Glu)
GDP-4-dehydro-6-deoxy-D-mannose
2-Phenyl-1,3-propanediol
monocarbamate
3-Carbamoyl-2-phenylpropionaldehyde
5-Phenyl-1,3-oxazinane-2,4-dione
Alcophosphamide
Aldophosphamide
GDP-L-fucose
D-Fructuronate
5,10-Methylenetetrahydrofolate
Xanthosine
5'-phosphate
6-Thioxanthine
5'-monophosphate
Adenylated
molybdopterin
6-Thioinosine-5'-monophosphate
IMP
Cobamide
coenzyme
alpha-Ribazole
UDP-N-acetyl-D-galactosamine
3alpha,7alpha-Dihydroxy-5beta-cholestan-26-al
3alpha,7alpha,26-Trihydroxy-5beta-cholestane
Secondary alcohol
Glyoxylate
Glycolate
UDP-N-acetyl-alpha-D-glucosamine
5-(5-Phospho-D-ribosylaminoformimino)-1-(5-phosphoribosyl)-imidazole-4-carboxamide
N-(5'-Phospho-D-1'-ribulosylformimino)-5-amino-1-(5''-phospho-D-ribosyl)-4-imidazolecarboxamide
3,4-Dihydroxymandelaldehyde
Ketone
L-Glutamyl-tRNA(Glu)
4-Hydroxy-5-phenyltetrahydro-1,3-oxazin-2-one
3,4-Dihydroxyphenylethyleneglycol
D-Glucuronate
Precorrin 6Y
Betaine aldehyde
Molybdate
Molybdoenzyme
molybdenum
cofactor
trans-3-Chloroallyl
aldehyde
FADH2
FAD
Precorrin 6X
Thiocysteine
L-Cysteine
O-Succinyl-L-homoserine
L-Cystine
Anthranilate
NH3
O-Acetyl-L-serine
L-Citrulline
Thiosulfate
Iminoaspartate
N-Carbamoyl-L-aspartate
L-Glutamine S-Sulfo-L-cysteine
L-Selenocysteine
6-Thioguanosine
monophosphate
Nitrogen
Nitrite
Hydrogen
selenide
HCO3-
Sulfite
Carbamoyl
phosphate
GMP
Selenite
Glycerone
phosphate
1-(2-Carboxyphenylamino)-1-deoxy-D-ribulose
5-phosphate
O-Phosphorylhomoserine
Porphobilinogen
Nicotinate
D-ribonucleotide
N-(5-Phospho-D-ribosyl)anthranilate
5-O-(1-Carboxyvinyl)-3-phosphoshikimate
D-Glyceraldehyde
3-phosphate
D-Erythrose
4-phosphate
beta-D-Fructose
6-phosphate
beta-D-Fructose
1,6-bisphosphate
Hydroxymethylbilane
Sedoheptulose
7-phosphate
Precorrin 1
Methanethiol
Uroporphyrinogen
III
S-Adenosyl-L-methionine
L-Methionine
Mannitol
D-Mannitol
1-phosphate
L-Methionyl-tRNA
UDP-glucose
3-Dehydroquinate
Orthophosphate
Indoleglycerol
phosphate
3-Dehydroshikimate
N-Acetyl-L-glutamate
5-semialdehyde
Shikimate
N-Acetyl-L-glutamate
5-phosphate
Indole
Shikimate
3-phosphate
Sedoheptulose
1,7-bisphosphate
D-Glucose
1-phosphate
Folinic acid
5-Phospho-alpha-D-ribose
1-diphosphate
ATP
Cytidine
L-Tryptophan
Uridine
1-(5-Phospho-D-ribosyl)-ATP
1-Hydroxymethylnaphthalene
dTDP-galactose
dTDP-glucose
5,10-Methenyltetrahydrofolate
Starch
1-Octanol
1-Octanal
UMP
tRNA(Trp)
ADP-glucose
L-Tryptophanyl-tRNA(Trp)
D-Fructose
1,6-bisphosphate
Figure 3.11: Continous Experiment, Solventogenesis, with Augmentation, green
and red: Hconti(s1), red: Gconti(s1)
3.3. DATA-DRIVEN PATHWAYS 53
3.3.6 Conclusions
In this section several data were presented with the previously derived formalism
for data-driven pathway generation. As this model was used for visualisation only
purposes, it was not curated. As soon as the download from published curated
databases is possible, visualisation of these pathways can be re-done easily because
of script automation (appendix B.2).
Biological Activity is Critical for the Choice of the Boundary Parameter
As first challenge, a scheme was researched at which transcript expression level
a reaction is considered active in the network. The boundary parameter
b
was
derived using a general graph topological trait, the edges to nodes fraction in
combination to general graph statistics. Similar proceeding are known to be fruitful
in pathway recognition [
Khatri et al., 2012
]. The precise network and its outcome
still heavily depend on the choice of
b
. While its definition is intuitive, it includes
the assumption that it is single-valued for each experiment. Transcriptional
activity, defined as the activity of RNA polymerase, is known to differ along
the bacterial life cycle [
Golding et al., 2005
]. Consequently, each state could also
obtain its individual
b
accounting for different transcriptional activities. On the
one hand, the design of the reference state in the microarray experiment copes
with that problem: By taking the average over all time-points as reference, the
variations in transcriptional activity can be accurately covered. On the other
hand, other data reference to a state at the beginning or the end of cultivation,
and thereby they distort the data. Time-dependent choice of
b
is comparable to a
segmented normalisation of data, proven superior before [Yang et al., 2003].
Improved Evaluation of the Boundary Parameter by Qualitative Biological
Knowledge?
Statistical tools to assess significant changes in expression pattern were used
[
Cakir et al., 2006
], this is comparable to the determination of a significant
b
.
It is equally reported that statistical tools work on the edge of their intended
capability, non-statistical tools are proven more worthy [
Huang et al., 2009
]. An
optimal parameter for the transcriptome evaluation can be found by an alternative
route - complementation with pathway information. Expectations on compound
connectivity are biologically testable, e.g. the number of reactions converting
ATP or NADH is one feasible criterion.
Alternatively, changes in product concentrations can be mapped to changes in
respective transcripts. Accumulation of metabolite pools suggest that there are
more influx than outflux reactions. This naturally requires a better curated
network.
From a comparative perspective, one can also note that the transcriptome data
54 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
can be split into two functionally different sets, a set of enzyme-coding transcripts
and a set of non-enzyme coding transcripts as will be shown in the next section
for a different purpose (equation 3.21). Both subsets are correlated because they
represent the vital organism. If it were possible to access all transcriptional
regulators in C. acetobutylicum as it is for E.coli [
Gama-Castro et al., 2008
] with
their respective open reading frames (ORFs), deduction of
b
requires that these
sets correlate for a given b.
Augmentation Requires More Extensive Studies
After the proposition of the augmentation rule (section 3.2.4) its application showed
that it acts very differently on the solventogenic state than on the acidogenic
state - while in acidogenesis an influence could be noted, solventogenesis did not
show alternated graph topologies.
This result relies on the sizes of the three different regions that are constructed
by choosing the boundary parameter
b
. This choice is critical at the point where
genes vary close to the boundary δb
δb=|x(s1)−x(s2)|= (−b−ǫ−(−b+ǫ)) = 2ǫ≈0
Such genes are falsely considered as augmented: In continuous culture 24% of all
augmented genes from acidogenesis differ by less than one order of magnitude,
during solventogenesis that is 20%.
It is suggested to further investigate other augmentation rules, e.g.
X0
b
(
s
) that
circumvents this problem.
Multidimensional Visualisation is a Challenge
Multi-dimensionality easily arises in the biological context. Here, the visualisation
of KEGG database was undertaken by a metabolite-metabolite mapping with
integrated data. The underlying structure is however more complicated because re-
actions usually require more than one substrate to produce more than one product.
Such graphs are hypergraphs; sets of nodes are connected by sets of edges. The
visualisation of such problems is only at its beginnings [Junghans, 2008].
Already for the simple graphs shown here visualisation is a challenge in Cytoscape.
It was the aim to facilitate hypothesis generation. In particular, visualisation
of ontologies for genes, enzymes, reactions, metabolites and pathways require
side-by-side visibility. Several graphs and their different attributes were shown, the
two boundary parameters in the batch experiment, the two types of logical rules
in the continuous experiment, the network centrality measures in the stimulated
batch experiment. More annotation needs to be fed to the graph in order to
allow easier interpretation, up to five dimensions are possible in one graph, still
Cytoscape is not equipped for such a purpose. Enrichment tools are frequently
3.3. DATA-DRIVEN PATHWAYS 55
encountered [
Huang et al., 2009
] and efforts are spent in visualising ontologies
in web-interfaces [
Dennis et al., 2003
], in trees [
Chevenet et al., 2006
], a tool for
multidimensional annotation visualisation in graphs seems missing so far.
Dynamic visualisation by movies shows the emergence of paths and their disap-
pearance. Usability assessments of this approach indicate that static networks are
easier to treat [
Farrugia and Quigley, 2011
]. Dynamic and static possibilities to
visualise different timescales are reviewed by [Secrier and Schneider, 2013].
Use of this Model
From the networks of batch culture and stimulated batch culture it was shown
that a state comparable to solventogenesis was induced by acetate addition. This
view is supported by experiments in continuous culture according to COSMIC
specifications in which acetate stimuli (50
mM
) as pulse or as step-function were
applied during acidogenesis. The acetate was used for the production of acetone
and butanol. It was also found in these cultures that glucose uptake is stimulated
and growth starts.
It was further shown that positions of metabolites within the data-driven network
can be monitored across different metabolic states and pathways. The appearance
of crotonate indicates a pathway that contains this metabolite. It will be shown
in the next section that the observance of this metabolite leads to the formulation
of a new pathway and the annotation of yet unannotated proteins.
56 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
3.4 Identification of Missing Reactions
Missing annotations for reactions are frequently encountered. One possibility to
deal with these gaps is to transfer annotation from other species [
Forslund, 2011
].
This section lays the fundamentals to compare organisms based on a KEGG-
database query and Pfam-motifs (3.4.1). Comparing a close relative to C. acet-
obutylicum which is B. subtilis (3.4.2) is carried out and hypotheses can be
constructed side to side to the MMM.
3.4.1 Comparative Approach
Motivation
The original annotation of C. acetobutylicum is incomplete, a large set of genes
has no annotation. Recently reported experiments suggest that there may be
branches in the acid and solventogenic pathways missing, because a knock-out
of the transacetylase and transbutyrylase still yielded an acetate and butyr-
ate positive mutant [
Lehmann et al., 2012a
,
Lehmann et al., 2012b
]. Also the
hitherto unannotated tricarboxylic acid-cycle was only recently discovered by
a metabolome study [
Crown et al., 2011
]. Determination of function is known
to be possible through homologies in close relatives [
Durot et al., 2009
]. The
use of Pfam-motifs [
Punta et al., 2012
] and Pfam-motif architecture has therefor
gained increasing interest during the last years [
Ofran et al., 2005
,
Lin et al., 2006
,
Koestler et al., 2010].
Database Query of Missing Reactions
Assume there is a second organism to which one can compare the clostridial
reactome. Applying the following database query efficiently identifies missing
reactions in one organism by comparing functional similar homologues in the
other. First, the reactome and the occurring Pfam-motifs will be harvested. As
before the set of reactions an enzyme can perform (
RX
), is given by a map
Rct
from a specific set of genes
X ⊂ XKEGG
which forms a subset of all genes in the
database, here the genes in KEGG XKEGG.
RX:= RctX.(3.18)
A similar map
Pfa
is given to determine the constituting Pfam-motifs for each
protein from the gene sequence (PX):
PX:= PfaX.(3.19)
It will be required to determine the inverse of the map
Rct
. For a specific reaction
r∈RXit is given by:
Rct−1(r∈RX) := nx∈ X KEGG :r=Rctx)o.(3.20)
3.4. IDENTIFICATION OF MISSING REACTIONS 57
From this definition it becomes clear that the inverse takes values in the whole set
of genes within KEGG. The inverse of a specific Pfam-motif is defined similarly.
Not every protein is bearing catalytic functions, therefore the kernel
X0
of
Rct
is
useful to enhance computational efficiency:
RctX0=∅.(3.21)
This kernel partitions the set of genes accordingly to their existing reaction-
annotation XReact.
XReact := X \ X0(3.22)
Pfam-Motif Comparison Between Two Organisms
Choosing a close relative of C. acetobutylicum, e.g. B. subtilis, one can assume
that the reactome of B. subtilis is better curated than the one of Clostridium.
In order to compare both reactomes, one first determines reactions specific to B.
subtilis (RBS
spec) by subtracting common reactions from both reactomes:
RBS
spec := RctXBS
React\RctXCA
React(3.23)
The inverse of this map now shows all the genes specific to these unknown reactions
in Clostridium but known in Bacillus. Intersection with the genes of Bacillus
narrows the solution space to the genes of interest (XBS
spec) in this study:
XBS
spec := Rct−1RBS
spec∩ X BS (3.24)
The Pfam-motifs of interest
P
are now determined from this set of specific
genes
PfaXBS
spec
and retrieved in the genes with no reaction annotation of
Clostridium
PXCA
0
. The two set of genes (
Xcomp
) bearing at least one of
these motifs then will serve as database for comparison of these two species:
P:= PfaXBS
spec∩PXCA
0(3.25)
XCA
comp := Pfa−1P∩ XCA (3.26)
XBS
comp := XBS
spec (3.27)
Similarity Measure for Functional Homologs
Connection of the two gene sets
XCA
comp
and
XBS
comp
according to their Pfam-motifs
Pyields a map.
This map however requires reduction to be useful for manual inspection: Sim-
ilarity in one motif only is not sufficient to hypothesise on the same function.
Consequently, it is necessary to restrain the connections by assuming, that two
genes of Bacillus and Clostridium are connected only when at least a percentage of
their motifs is similar. Established similarity measures for Pfam-motif comparison
58 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
are given in [
Lin et al., 2006
]. The Jaccard-index
J
(
x, ˆx
),
x∈ XCA,ˆx∈ X BS
[Levandovski and D, 1971] is one of them:
J(x, ˆx) = PCA(x)∩ PBS(ˆx)
PCA(x)∪ PBS(ˆx)(3.28)
This study proposes to use a different measure
nP
(
x, ˆx
) for such purpose as will
become clear instantly:
nCA
P(x, ˆx) := PCA(x)∩ PBS(ˆx)
|PCA(x)|(3.29)
nBS
P(x, ˆx) := PBS(ˆx)∩ PCA(x)
|PBS(ˆx)|(3.30)
nP(x, ˆx) := 1
2nCA
P(x, ˆx) + nBS
P(x, ˆx).(3.31)
Comparison of Similarity Measures
Consider the three following general cases of two sets P(g1) and P(g2):
1. case: P(g1) = {P1, P2}versus P(g2) = {P1,· · · , P10}
2. case: P(g1) = {P1,· · · , P4}versus P(g2) = {P3,· · · , P12}
3. case: P(g1) = {P1,· · · , P4}versus P(g2) = {P3,· · · , P6}
Obviously for case 1,
P
(
g1
) contains all items that are present in
P
(
g2
). It is
possible that
P
(
g2
) functions as
P
(
g1
), only
P
(
g1
) has more motifs. The Jaccard-
index
JP
(
g1
)
, P
(
g2
)
= 0
.
2 is less beneficial in this case then the here proposed
similarity measure nP(g1), P(g2)= 0.5(1 + 0.2) = 0.6.
The overlap in case 2 is identical to case 1, now
P
(
g1
) contains two distinct
motifs to
P
(
g2
), the sizes of proteins are as in case 1. Again, the Jaccard-index
JP(g1), P(g2)=1
6is less beneficial nP(g1), P(g2)= 0.5(0.5 + 0.2) = 0.35.
Case 3 studies the effect if the sizes of both proteins are equal, both indices are
increased to
JP
(
g1
)
, P
(
g2
)
= 0
.
33
¯
3
and
nP
(
g1
)
, P
(
g2
)
= 0
.
5(
2
4
+
2
4
) = 0
.
5.
The Jaccard index indicates again a very low similarity, although half of the
motifs in both proteins are equal. These three examples show that the Pfam-motif
similarity
nP
gives a bonus to small proteins. If their function is known annotation
transfer to larger proteins is enhanced.
3.4.2 Comparison of B. subtilis and C. acetobutylicum
The calculated sets from the previous presented approach are listed in table 3.4.2.
The number of genes with a reaction annotation is higher for Bacillus than for
Clostridium (747 vs 600). Consequently, the number of reaction specific to Bacillus
is more than twice as large as the number of reactions in Clostridium (388 vs 139).
3.4. IDENTIFICATION OF MISSING REACTIONS 59
Table 3.3: The download of information from KEGG concerning both organisms
resulted in approximately equal sized genomes (
X
). The Bacillus genome contains
more genes with reaction-annotation (
X\ X0
) and more reactions (
RX
) than the
one of Clostridium. There are twice as much reactions that can be inferred from
B. subtilis than from C. acetobutylicum (
Rspec
) for the respective other organism.
267 genes are responsible for these reaction (
Xspec
). 512 motifs are found in this
specific gene-set of Bacillus.
sets B. subtilis C. acetobutylicum
size size
X4422 4021
X \ X0747 600
RX1041 792
PX5004 4494
Rspec 388 139
Xspec 267 145
Xfinal 204 821
A map of
RBS
spec
is given in figure 3.12. This map is one tool to track reactions
that are not present in C. acetobutylicum. Connection of the genes according
to their Pfam-motifs yields a second tool for rapid function suggestion (figure
3.13). Direct annotation transfer is possible for smaller connected components
in this unfiltered map by considering high edge-weights because several protein
functions are contained in few very specific Pfam-motifs, e.g. CoA-transferase
activity. The correct reaction-annotation is given for several enzymes (table 3.4.2),
however the protein does not contain an E.C. number in C. acetobutylicum. This
proof of concept shows also that hypothetical proteins with unknown functions
are mapped to proteins with functional annotation, suggesting hypotheses can be
derived from such maps.
Possible model reduction is achieved by choosing a threshold for the edges weights
filtering all edges with nP<0.5, reduces the model size by 66% (figure 3.14).
60 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
Figure 3.12: The reduced MMM of B. subtilis shows the reactions that are not
annotated in C. acetobutylicum. As before, the size of the nodes correlates to
stress and the colour to the eccentricity, red correspond to high values, and green
to low values. This network consists of 577 metabolites, thereof 334 unique to B.
subtilis.
rectangle: the compound is present only in B. subtilis.
ellipsoid: the compound when is present in both organisms.
3.4. IDENTIFICATION OF MISSING REACTIONS 61
Figure 3.13: Genes of C. acetobutylicum are connected to the genes of B. subtilis
which have an reaction annotation that is not found in C. acetobutylicum. The size
of the nodes correlates with stress, and the colour corresponds to the eccentricity
on an increasing green to red scale.
62 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
Table 3.4: Small connected components of this mapping either reproduce, expand
or induce functions on proteins. The two columns of genes are connected to each
other in the database. The first block of this table shows examples of enzymes that
are correctly annotated and mapped together. These genes in Clostridium are
not completely annotated as enzymes and represent gaps in the KEGG database.
The second block represents the genes with poor or missing annotation, here the
annotation can be considerably enriched with the Bacillus information.
DH = dehydrogenase, ST = sulfotransferase
xAnnotation ˆxAnnotation
CAC0997 nucleoside-diphosphate kinase BSU22730 same (EC:2.7.4.6)
CAC1200 phospho-adenylylsulfate ST BSU10930 same (EC:1.8.4.8)
CAC1200 phospho-adenylylsulfate ST BSU15570 same (EC:1.8.4.8)
CAC1462 levanase/invertase BSU34460 same (EC:3.2.1.65)
CAC1462 levanase/invertase BSU27030 same (EC:3.2.1.65)
CAC3498 ribokinase sugar kinase BSU35920 same (EC:2.7.1.15)
CAC1574 4-hydroxybutyrate DH BSU31050 choline DH (EC:1.1.1.-)
CAC3392 butanol DH BSU31050 choline DH (EC:1.1.1.-)
CAP0059 alcohol DH BSU31050 choline DH (EC:1.1.1.-)
CAC0804 Pectate lyase related protein BSU07560 pectate lyase (EC:4.2.2.2)
CAC1190 Fe-S-cluster redox protein BSU32330 lipoyl synthase [EC:2.8.1.8]
CAC1229 hypothetical protein BSU10250 lipoate-protein ligase (EC:6.3.2.-)
CAC3238 hypothetical protein BSU32330 lipoyl synthase [EC:2.8.1.8]
3.4. IDENTIFICATION OF MISSING REACTIONS 63
Figure 3.14: The distribution of edge weights of the comparative approach.
64 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
3.5 Annotation Transfer a Case-Study: 3-HBDH Activity
in Clostridium
Here the example research of an enzyme activity in C. acetobutylicum will be
carried out. After introduction of the general approach (3.5.1), experimental
indications and database information are collected (3.5.2 and 3.5.3), then hypo-
theses will be given (3.5.4) and methods results will be shown (3.5.5). Finally
some conclusions will be drawn (3.5.6).
3.5.1 Annotation Transfer Methods
Available Annotation Transfer Methods
Three different methods (
M1, M2, M3
) will be considered for annotation transfer.
M1:
BLASTP mapping [
Altschul et al., 1997
] of two protein sequences, e.g. a
protein of C. acetobutylicum to a close relative like B. subtilius was repor-
ted to detect enzymes with a comparable function [
Rost et al., 2003
] and
architecture [Lee et al., 2008a].
M2:
Phylogenetic approaches have been used with success for annotation transfer
[
Pellegrini et al., 1999
]. Comparing the Pfam-motifs of a chosen protein
throughout different species gives a second approach. Indeed, combinations
of domains are enriched in some functional classes [
Forslund, 2011
]. Pfam-
domain architecture is accessible by several tools, e.g. the Weighted Domain
Architecture Comparison Tool (WDAC) [
Lee and Lee, 2009
] or the Feature
Architecture Comparison Tool (FACT) [
Koestler et al., 2010
]. Here, the
Pfam-motifs responsible for an enzymatic activity in other organisms will
be determined from a frequentist point of view and then these motifs will
be retrieved in the clostridial annotation. This can be considered a pre-
selection step before the more exhaustive Pfam-motif architecture approach
is calculated.
M3:
Enzymes of one pathway are known to build stoichiometric complexes that
channel the substrate [
Srere, 1987
]. Clustering of gene expression data
is frequently used to reveal open reading frames and co-regulated genes
[
Tavazoie et al., 1999
,
Dhaeseleer et al., 2000
]. By fixing a regulatory as-
sumption, e.g. by choosing a gene of a target pathway, all possibly co-
regulated genes are identified.
3.5. ANNOTATION TRANSFER A CASE-STUDY: 3-HBDH ACTIVITY IN
CLOSTRIDIUM 65
Integration of Annotation Transfer Methods
Any annotation transfer method
M
should be able to create subsets from the
whole set of genes, they are named candidates XMi
XMi:= Mi(X).(3.32)
In order to evaluate the retrieval of a candidate by a method, a score-matrix (
CM
)
is defined with the NMmethods as columns and the NJgenes as rows:
CM:= (1; x∈ XMi
0; x /∈ XMi
(3.33)
Combining these counters in a ranking
sM
with a column-vector of weights
w
enables the integration of all methods to one numeric value.
sM:= CM
w
||w|| (3.34)
3.5.2 Collection of Experimental Indications
Indications for the Presence of Crotonate
It was mentioned earlier that during the batch fermentation data a delta-2
oxidoreductase is up-regulated during acidogenesis and down-regulated during
solventogenesis (3.3.3). The same pattern is obvious in the two stationary states
of the continuous culture (figure 3.9 and figure 3.11). Measurements confirm that
indeed crotonate is present in small amounts during continuous culture
∗
. This
suggests that there may be a pathway that uses crotonate.
Indications for a Unreckoned Butyrate Production Pathway
Mutants of acetate kinase or butyrate kinase did always produce minor amounts
of both acids [
Green et al., 1996
] and butyrate was taken up by an unknown path-
way - the established reverse kinase pathway and CoA-transferase activity were
knocked-out [
Lehmann et al., 2012b
]. An acetoacetate decarboxylase knock-out
mutant produces increased amounts of butyrate when supplemented with calcium
carbonate and methyl viologen [
Jiang et al., 2009
]. In a different culture, knock-
out mutants of acetoacetate decarboxylase have increased butyrate concentration
and acetoacetate does not accumulate [
Lehmann et al., 2012a
]. Acetoacetate addi-
tion in pH-uncontrolled culture leads to increased butanol and butyrate production.
In pH-controlled culture this effect is less pronounced [
Papoutsakis et al., 1987
].
This suggests that butyrate may be processed back to acetoacetate and vice versa
under particular circumstances.
∗personal communication, Kengen Laboratory, Wageningen
66 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
3.5.3 Collection of Database Information
Simplistic Approach to Alternative Production Pathways
Investigation of the clostridial reactome shows that no other reaction than the
delta-2 oxidoreductase (E.C. 1.3.1.31, rn:R01689) is able to use crotonate as
substrate or product. It seems unreasonable that an enzyme exists without any
further integration of its substrates or products in other pathways. A simplistic
approach would acknowledge that the enzymes (CA
C2708
,CA
C2710
,CA
C2711
)
using the CoA-derivates acetoacetyl-CoA, crotonyl-CoA, 3-hydroxybutyryl-CoA
have a broad specificity towards non-CoA derivates. However, this unspecific
reaction is not annotated in KEGG for these enzymes, also the crotonase is not
annotated to process CoA-derivatives. It is further known that C. acetobutylicum
possesses several distinct enzymes that do the same reactions, e.g. butanol
dehydrogenases [Grimmler et al., 2011], [Duerre, 2005, p.678].
Complex Approach to Alternative Production Pathways
Plausible other possibilities of this alternative pathway are presented in figure
3.15. Variant A assumes that other CoA-transferases than ctfAB are producing
the corresponding acids and CoA-derivates. Variant B assumes an unannotated
dehydrogenase in C. acetobutylicum that uses acetoacetate as substrate and
produces 3-hydroxybutyrate (B1) which is then the substrate for an unannotated
dehydratase to produce crotonic acid (B2). Variant C suggest that this pathway
can be inversed in direction.
There are No Unknown CoA-Transferases
The determination of variant A is straightforward through Pfam-motifs. The
motif for a general CoA-transferases is the CoA trans-motif. Which is available
in C. acetobutylicum only in the already mentioned proteins ctfAB (CA
P0163
and
CA
P0164
). This finding is complemented by a transacetylase inactivation strain
that was not able to re-assimilated acids [
Green et al., 1996
]. Substrate specificity
of the CoA-tranferases is broad, crotonate may be used as substrate, with an
activity loss of 39% [
Hartmanis et al., 1984
]. Other authors report various other
CoA-thioester substrates and cofactors for this enzyme [Barker et al., 1978].
No 3-Hydroxybutyrate Dehydrogenase is Known in C. acetobutylicum
Variant B1 is performed by the 3-hydroxybutyrate dehydrogenase (3-HBDH, EC
1.1.1.30, rn:R01361). Interestingly, no reaction for variant B1 is found in C.
acetobutylicum, the only way metabolising acetoacetate is via decarboxylation
[
Papoutsakis et al., 1987
]. However, in 750 other organisms there is this activity,
3.5. ANNOTATION TRANSFER A CASE-STUDY: 3-HBDH ACTIVITY IN
CLOSTRIDIUM 67
Figure 3.15: Alternative models of butyric ccid production suggest three different
variants of crotonate production via CoA-transferases (variant A), acetoacetate
consumption (variant B) or the inversed pathway (variant C)
one of them is B. subtilis (
BSU38970
, yxjF). By considering the two established
tools for comparison of C. acetobutylicum to B. subtilis the 3-HBDH reaction is
visible in the metabolic network (figure 3.12) and several proteins similar to the
Bacillus 3-HBDH are found (figure 3.13).
No 3-Hydroxybutyrate Dehydratase is Known
Finally, variant B2 is a dehydratase activity. While a motif is reported for CoA-
dependent dehydratases, there is no known reaction catalysing the conversion
from 3-hydroxybutyrate to crotonate. This reaction can only be an unspecific
byproduct of another enzyme, or a product of a multi-step reaction within one
enzyme. Fortunately, Pfam-motifs have further annotations, e.g. the Epimerase-
motif which has a hydratase annotation.
3.5.4 Hypotheses for Annotation Transfer
M1:
BLASTP of the clostridial proteome against yxjF (BSU
38970
) using KEGG.
M2:
A phylogenetic comparison of Pfam-motifs from all 750 annotated 3-HBDH
to the clostridial genome using Taverna and MATLAB (B.1 and B.2).
68 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
M3:
Clustering of batch (ba) [
Jones et al., 2008
] and continuous (cu) culture data
[Grimmler et al., 2011] according to three different regulatory scenarios:
A1:
Expression of the 3-HBDH is co-regulated to genes of the pathway
that converts acetoacetyl-CoA to butyryl-CoA under acidogenic condi-
tions, the responsible enzymes are encoded in the transcripts CA
C2709
,
CAC2711, CAC2712 [Jones et al., 2008,Grimmler et al., 2011] (ac).
A2:
Expression of the 3-HBDH is co-regulated to genes of the pathway
that converts acetoacetyl-CoA to butyryl-CoA under solventogenic
conditions, the responsible enzymes are encoded in the transcripts
CA
C2009
, CA
C2012
, CA
C2016
[
Jones et al., 2008
,
Grimmler et al., 2011
]
(so)
A3:
Expression of the 3-HBDH is co-regulated to the gene coding for the
enoate-reductase CAC3371 (er).
Clustering is performed by Genesis, [
Sturn et al., 2002
] using a kmeans
algorithm and the euclidean distance metric.
3.5.5 Results
Results of the BLASTP Approach
The gene BSU
38970
from B. subtilis was used as example for an annotated 3-HBDH
in a close relative. The BLASTP matches are noted in table 3.5. A small E-value
shows significant matches to the target structure. However, the size of the protein
may be significantly smaller than the target structure.
Results of the Phylogenetic Approach
750 organisms contain annotated 3-HBDH, harvest of their Pfam-motif showed
that the average value of motif per enzyme is nine. Within the frequent motifs
(table 3.6) the two adh-motifs and the KR-motif are predominant. They could
serve as first criterion for research of the 3-HBDH. Following the list, several
NAD-motifs are preserved throughout most of the species. One further observes
the 3HCDH N-motif and the Epimerase motif as being characteristic.
With the adh short and the adh short C2 motif, 749 organisms are already covered.
The only organism not covered is Brucella melitensis, its 3-HBDH contains
only low abundant motifs: the TrkA N, the 3HCDH N and the 2-Hacid dh C.
Eleven genes contain only the adh short C2-motif, they are found in the genus
Rickkettsia. Another eleven organisms contain only the adh short-motif. These are
more heterogeneously spread than for the adh short C2-motif, containing higher
vertrebrates (Pongo abelii or Sus scrofa) and bacteria, e.g. Desulfobacterium
autotrophicum.
3.5. ANNOTATION TRANSFER A CASE-STUDY: 3-HBDH ACTIVITY IN
CLOSTRIDIUM 69
Table 3.5: BLASTP of the aminoacid sequence of
BSU38970
to the proteome of C.
acetobutylicum
Gene-ID E-value
CAC2607 5.00E-38
CAC3574 5.00E-37
CAC3462 9.00E-28
CAC0361 3.00E-27
CAC2626 3.00E-23
CAC1423 7.00E-17
CAC3335 9.00E-17
CAC1576 5.00E-15
CAC2992 8.00E-15
CAC1331 1.00E-13
CAC0536 4.00E-10
CAP0051 6.00E-09
CAC3484 5.00E-08
CAP0001 7.00E-07
CAC3355 8.00E-04
The number of considered motifs determines the number of candidates. From the
pure frequentist point of view, the first four motifs seem promising as model for
the 3-HBDH activity if not the Brucella gene would be annotated with the same
function but without these motifs. In order to increase the list of candidates also
low abundant motifs (at least 10% matching) will be considered as relevant. The
corresponding solution set contains 79 proteins.
Results of the Regulatory Approach
The four genes CA
C2708
,CA
C2710
,CA
C2711
,CA
C3371
from the regulatory assump-
tions, were located in three different clusters during batch and continuous culture
(refer to figure 3.16). For the operon that is up-regulated under solventogenic
conditions (CA
C2009
,CA
C2012
,CA
C2016
) the same cluster partition was used. Nat-
urally within the results, all genes from the assumptions occur in these clusters.
Three other genes are co-regulated in both experiments (
ba
,
co
) and bearing
similar motifs: CA
C2713
, the redox-sensing transcription repressor Rex, which is
in the same open reading frame, a ketopantoate reductase named PanE/ApbA,
CAC2937 and a nucleoside-diphosphate-sugar epimerase, CAC2166.
The ketopantoate reductase, shares several motifs with CA
C2708
: a NAD binding 2-
motif, three different dehydrogenases, the NAD Gly3P dh N-motif, the 3HCDH N-
motif, responsible for the reduction of 3-hydroxyacyl-CoA and NAD-binding and
the UDPG MGDP dh N-motif. Finally, the ApbA-motif responsible for keto-
pantoate reductase activity is found in both proteins. Domains unique to the keto-
70 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
Table 3.6: List of 3-HBDH motifs most frequently occurring in 750 annotated
KEGG species.
Motif Frequency
adh short 0.98
adh short C2 0.98
KR 0.97
Epimerase 0.70
Eno-Rase NADH b 0.42
3HCDH N 0.40
NAD binding 10 0.35
Polysacc synt 2 0.23
Saccharop dh 0.21
DUF1776 0.19
TrkA N 0.16
THF DHG CYH C 0.15
3Beta HSD 0.14
RmlD sub bind 0.13
AdoHcyase NAD 0.13
Shikimate DH 0.11
2-Hacid dh C 0.10
pantoate reductase are a synthase for heptaprenyl diphosphate HEPPP synt 1,
DUF1879, a domain of unknown function, and a C-terminal ApbA-motif.
It should be noted, that CA
C3371
shares many similar functions to CA
C2708
as well,
suggesting that this gene may also have more activities than already annotated.
Integration
The precedent approaches, the BLASTP-search, the phylogenetic comparison
and the clustering all show different candidates (3.7). These weights are chosen
to add to 9 for the experimental and for the database assisted methods. Since
a gene can only be contained within one cluster for either culture method, the
maximal score achievable is 6. The only exception to this occurs when two
assumptions are contained within the same cluster, e.g. the CAC
3335
is contained
in the clusters of CA
C2708
, CA
C2709
, CA
C2711
in batch culture, and the A1 and
A3 condition partially overlap for three candidates. The score further prefers
the database methods to the experimental methods by three points (9 versus
6 for batch and continuous culture). BLASTP is considered much inferior to
Pfam-motif search (2 versus 7). The final ranking is shown side-to-side with
the Pfam-motif similarity
nP
in table 3.8. It is obvious from the table, that
integration of different methods drastically changes the ranking compared to the
Pfam-motif frequency alone. The sugar epimerase CA
C2166
, the ketopantoate
72 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
Figure 3.17: Map of 3-HBDH candidate genes and their Pfam-motifs.
green: gene identifier, red: motifs not retrieved by the phylogenetic Pfam-research,
darkblue to white: increasing frequency of occurrence in the phylogenetic Pfam-
research
3.5. ANNOTATION TRANSFER A CASE-STUDY: 3-HBDH ACTIVITY IN
CLOSTRIDIUM 73
Table 3.7: Integration of the three approaches with the assumptions and different
datasets results in a data matrix Cthat is truncated to high score genes.
weights w 2 7 3 3 3 3 3 3
Genes M1M2M3ba,ac M3ba,so M3ba,er M3co,ac M3co,so M3co,er
CAC2166 0 1 1 0 1 1 0 0
CAC3371 0 1 1 0 1 0 0 1
CAC0267 0 1 1 0 1 0 0 0
CAC2009 0 1 0 1 0 0 1 0
CAC2708 0 1 1 0 0 1 0 0
CAC2713 0 1 1 0 0 1 0 0
CAC2937 0 1 1 0 0 0 0 1
CAC3335 1 1 0 0 0 0 0 1
CAC3355 1 1 0 1 0 0 0 0
Table 3.8: Application of the defined numerical weights produced priorities
the candidates, the concomitant mapping of similarity to the B. subtilis Pfam-
annotation represents a second dimension of information.
Genes gAnnotation sMnP(g, BSU38970)
CAC2166 nucleoside-diphosphate-sugar epimerase 0.59 0.65
CAC3371 2-enoate reductase 0.59 0.06
CAC0267 L-lactate dehydrogenase 0.48 0.23
CAC2009 3-hydroxyacyl-CoA dehydrogenase 0.48 0.28
CAC2708 3-hydroxybutyryl-CoA dehydrogenase 0.48 0.25
CAC2713 redox-sensing transcriptional repressor Rex 0.48 0.07
CAC2937 ketopantoate reductase PanE/ApbA 0.48 0.34
CAC3335 Short-chain alcohol dehydrogenase enzyme 0.44 0.63
CAC3355 polyketide synthase 0.44 0.21
each other. If there is a 3-HBDH in C. acetobutylicum experiments should aim at
this gene first and then the two others.
3.5.6 Critical Evaluation
Starting from the research of unannotated reactions, a tool was constructed
that compared two organisms based on their statistical Pfam-motif similarity.
Differences of annotated reactions are readily visualised by a second tool, the
mapping of
Rspec
in a MMM. From these two tools and experimental data,
an alternative production pathway of butyrate was proposed. The researched
3-HBDH activity was conducted by integration of three different methods for
annotation transfer, thereof one established, the BLASTP, one current approach
in its simpler version, the Pfam-motif search, and regulatory investigation to
74 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
integrate experimental data by clustering. As a result, candidates for the 3-HBDH
in Clostridium acetobutylicum were proposed.
Remarks on Data
There is not yet strong evidence that a crotonate utilisation pathway is missing.
The deletion of the 3-hydroxybutyryl-CoA dehydrogenase results in a butyrate
and butanol deficient strain [
Lehmann and Luetke-Eversloh, 2011
]. This suggests
that at least one product of this enzyme has a fundamental role for this pathway,
e.g. crotonyl-CoA as CoA donor for hydroxybutyrate. Experiments with supple-
mentation of crotonic acid or acetoacetate in minimial medium in pH-uncontrolled
batch culture were inconclusive regarding the role of crotonate in the wildtype
(results not shown).
Remarks on Tools
The comparison of C. acetobutylicum and B. subtilis via their Pfam-motif similarity
nP
created a similarity map. This tool for the identification of annotation gaps
can be used for any other organism and also by employing different similarity
measures, as suggested by [
Lin et al., 2006
]. It is extendable by further organisms,
however visualisation then gets a bigger challenge. Also, the retrieval of a suitable
cut-off value for the similarity measure is not evident. Here an external criterion
must be found.
Remarks on BLASTP
BLASTP was earlier encouraged [
Rost et al., 2003
]. It assumes the two types of
enzyme are very similar, so it is required that yxjF is not a unique type enzyme as
e.g. the 3-HBDH of Brucella melitensis. The Pfam-motif map of candidates from
BLASTP is shown in figure 3.18. The 15 candidate proteins are centred around
mainly six motifs: adh short, adh short C2, KR, epimerase, NAD binding 10,
and 3HCDH N. As was shown in the phylogenetic approach, these motifs are
characteristic for most 3-HBDH and yxjF is indeed an adequate role-model for a 3-
HBDH activity. However, the use of BLASTP is now discouraged [
Forslund, 2011
]
because it is reckoned it is better suited to calculate a phylogenetic distance
than a functional distance. It is a comparative approach which can only identify
homologies, not analogies.
Remarks on Pfam-Motif Comparison
The here presented statistical approach for Pfam-motif comparison along different
phylogenies is only possible if the motif is unspecific. Targets like a CoA-transferase,
a fumarase, or a crotonyl-hydratase are readily identifiable by their respective
Pfam-motif. Still, the information content of two different Pfam-motifs is not
3.5. ANNOTATION TRANSFER A CASE-STUDY: 3-HBDH ACTIVITY IN
CLOSTRIDIUM 75
IATP
2-Hacid_dh_C
SEFIR
ZapA
DFP
Spore_III_AB
DivIC
Pyr_redox
RmlD_sub_bind
Eno-Rase_NADH_b
DUF1776
cac:CA_P0001
Epimerase
adh_short
NAD_binding_10
cac:CA_C2626
cac:CA_C2992
cac:CA_C3335
adh_short_C2
cac:CA_C1576
cac:CA_C3574
cac:CA_C2607
cac:CA_P0051
DUF4217
DUF1815
KR
cac:CA_C1423
PP-binding
ketoacyl-synt
Acyl_transf_1
Thiolase_N
Ketoacyl-synt_C
PA
PALP
cac:CA_C3355
cac:CA_C0361
NAD_binding_4
cac:CA_C3484
F420_oxidored
cac:CA_C3462
3HCDH_N
cac:CA_C0536
DUF2741 TH1
Polysacc_synt_2
DUF3608
NmrA
TrkA_N
Shikimate_DH
NAD_binding_7
3Beta_HSD
Oxidored_nitro
Saccharop_dh
cac:CA_C1331
Figure 3.18: Pfam-motif map of 15 BLASTP candidates with low E-values. KEGG
gene identifiers are marked in green hexagons, Pfam-motifs in orange circles.
identical and requires weighing according to their promiscuity along the kingdoms
[
Lee and Lee, 2009
]. For the phylogenetic comparison of 750 different species and
their 3-HBDH this was not urgently necessary because these enzymes are spread
through all kingdoms and only few domains were highly promiscuous, e.g. the
adh motif.
A sole statistic remains erroneous, for this reason online tools like WDAC and
FACT were used for the ultimate candidate selection. These tools require longer
calculation times and a full genomic research is a computationally demanding
task.
Pfam-motifs do not represent the sole possibility for annotation transfer, active-
site profiling has been proven successful for the identification of highly conserved
three dimensional structures of kinases from sequence data [
Cammer et al., 2003
].
76 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
Remarks on Clustering
Clustering has been proven worthy for functional annotation [
Dhaeseleer et al., 2000
].
Two plausibility criteria were used to restrain the candidate space. The first
assumes that the known pathway for acetoacetyl-CoA and acetoacetate reduction
runs in parallel to the unknown pathway. A similar approach was assumed by
[Brown et al., 2000] who related tumor proteins to ribosomal proteins.
Nevertheless, false negative candidates are a risk with clustering as with any other
data criterion since abstract assumptions are imposed: by choice of data curation
and standardisation, by choice of clustering algorithm, by choice of distance
metric [
Brown et al., 2000
,
Brohee and van Helden, 2006
,
Freeman et al., 2007
].
Despite this drawback, clustering induces a beneficial partition of the candidate
space and partition of the candidate space through clustering allows the step-
by-step elimination of invalid assumptions. If there is strong evidence that a
gene cannot be co-regulated to a member of a cluster, all other members are also
eliminated.
3.6. FINAL CONCLUSIONS 77
3.6 Final Conclusions
Overview
The manifold uses of pathway models were this chapter’s main topic: the contex-
tualisation of data, the guidance for metabolic engineering, the hypothesis-driven
discovery and the network property discovery [
Oberhardt et al., 2009
]. This
chapter started from the KEGG database and proposed Taverna as suitable tool
for the harvest of metabolite reaction networks, multidimensional annotation
and Pfam-motifs. For the analysis of the clostridial reactome, a formalism was
introduced that allowed the integration of transcriptome data and the formal
reaction-database to a data-supplemented database or a data-driven database. Sev-
eral visualisation softwares and visualisation methods were tested to allow manual
investigation of this database. From this, research of the 3-hydroxybutyrate
dehydrogenase activity within existing annotations was motivated. A scheme
was proposed to reveal candidates from integration of three different methods,
database related and experiment related.
Contextualisation of Data and Network Property Discovery
Huge amounts of data are produced and deposited, metabolomic data and tran-
scriptomic data stand side by side, the integration of omics is a necessary step
in research [
Joyce and Palsson, 2006
]. In 2011 the HITS-Institute published an
article that data-driven science represents a challenge for computer sciences and
a re-thinking of the roles of hypothesis driven approaches to organisation of
data into meaningful sets [
Reuter, 2011
]. Information retrieval and evaluation is
facilitated when data is partitioned into smaller sets [
Khatri et al., 2012
]. In this
thinking, the here presented pathway model approach represents one possibility
to coherently, self-consistently organise different data from different experiments:
Evaluation of transcriptome data on the reactome level reduces the number of
considered transcripts and it enables the evaluation the data in terms of graph
analysis.
The first step in this organisation is to make sense of the data and try then
to infer structures from the data. This is defined as a top-down approach by
[
van Riel, 2006
]. In several aspects, the here presented model contains a top-down
approach, because it integrated the data by using two logical rules, the result is a
subgraph that serves for several evaluations: Graph centralities help in manual
ranking metabolites according to their position in the network, the edges to
nodes fraction was able distinguish solventogenesis and acidogenesis, qualitative
assessments like the connectivity of a single metabolite through out the different
states provides first directions for its importance.
Known approaches of pathway analysis remain valid for this type of data-
base in order to create further metadata: Dynamic properties of the net-
78 CHAPTER 3. AUTOMATED NETWORK MODEL CREATION
work can still be derived [
Klipp et al., 2004
], network-motifs can be enriched
[
Joyce and Palsson, 2006
], elementary modes and cut sets can be calculated
[Klamt and Gilles, 2004].
Hypothesis-Driven Discovery and Guidance For Metabolic Engineering
The observation of crotonate synthesis is just one example how hypotheses are built
from such a model. Gene-Reaction Networks are frequently encountered in literat-
ure, still models concentrate on flux balance calculation [
Durot et al., 2009
], the al-
ternative evaluation of transcriptome data in the here proposed graph-based format
seems underrepresented for hypothesis finding and metabolic engineering strategies.
One reason for this is the focus on statistical evaluation of differentially expressed
genes [
Patil and Nielsen, 2005
]. Differential expression can be understood as a ma-
jority criterion for data reduction [
Yang et al., 2005
]. Within the spirit of person-
alised medicine and individual treatments [
W and BM, 2007
,
Katsnelson, 2013
],
the focus on single genes and their position within the whole network needs to be
restrengthened. This type of data-model should be understood as a complement
to statistical science. Crotonate synthesis is unreported in C. acetobutylicum,
what other annotations are missing in the published models?
The annotations of genomes is a long lasting process [
Khatri et al., 2012
].
This work proposes therefor a comparison tool and integration schemes of
data and the Pfam-database for annotation discovery. Within this scope
the here introduced metabolic networks, comparison tools and the ranking
score construct several what-if scenarios that aid in metabolic engineering
[
Aittokallio and Schwikowski, 2006
,
Durot et al., 2009
]. For several well studied
organisms a reactome knowledgebase is established [
Matthews et al., 2009
], this
work is a methodological contribution to it.
The experimental investigation of a hypothesis is the ultimate step. It seems
to be carried out with increasing ease: Multiple-site mutants of C. acetobutyl-
icum become more and more frequent, e.g. [
Jiang et al., 2009
,
Sillers et al., 2009
,
Lehmann et al., 2012a,Lehmann et al., 2012b].
Chapter 4
Automated Dynamic Model
Creation
Little by little, one travels far
John Ronald Reuel Tolkien
This chapter starts with the question which metabolic engineering strategies
can be employed in order to increase butanol production. A dynamic model of
butanol production that integrates time series of transcript data and metabolome
data will be therefor used. Existing approaches are reviewed first (4.1) and the
unique properties of this model introduced. From mass balance equations (4.2) a
formalism for the implementation in the IT-architecture will be given (4.3). This
model will be used for the parameter estimation of two different experiments (4.4).
Hypotheses will be generated by employing global sensitivity analysis (4.5) which
are followed by the conclusions (4.6).
80 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
4.1 Historical Perspective
Algebraic Rules For Flux Balance Analysis
Early studies on the stoichiometry of ABE-fermentations revealed simple al-
gebraic rules to connect data with unknown information like the energy value.
Their application was to check data inconsistencies [
Yerushalmi et al., 1983
,
Papoutsakis, 1984
]. These rules additionally lead to a program that enables
the use of constraint flux balance analysis (FBA) that gives insights in the flux dis-
tributions of the organism. It was used by [
Junne, 2010
] to evaluate the outcome of
stimulus response experiments in batch and continuous culture. Other applications
are known [Desai et al., 1999,Lee et al., 2008a,Senger and Papoutsakis, 2008].
Integration of mRNA Yielded Superior Results
The early model by [
Votruba et al., 1986
] is a data-based model that is driven by
curve fitting of different batch fermentation results to some function. Its kinetics
are Michaelis-Menten type kinetics or directly proportional with butanol based
inhibition terms. Direct relations between the different compounds through meta-
bolic pathways are not considered. However, this model introduces a metabolic
activity functional based on total RNA, which helps in describing the culture’s
history and consequently culture growth. In the same year, a model for glucose
uptake was published by [
Yerushalmi et al., 1986b
], they were proposing an active
site model that explains substrate internalisation and product externalisation.
The same authors also extended their model to relate mRNA concentrations and
butanol production from glucose. The mRNA is given the role to reflect culture
states. They assume diffusion of compounds through the cell membrane and
inhibition by butanol [Yerushalmi et al., 1986a,Yerushalmi et al., 1988].
Integration of Inhibitory Effects of Acids And Solvents
The model by [
Jarzebski et al., 1992
] aims at the understanding of a chemostat at
different pH values and thereby dissociation states of butyric acid. This coupling
influences growth and the onset of solventogenesis. A set of logical rules covers
inhibitory effects of butyric acid. This model describes the data well: Two steady
state values and a sustained oscillation are covered. Still, this type of logical
rules is insensitive to major system changes, as e.g. fermentations of mutants or
changes of medium composition. A product inhibition model was proposed by
[
¨
Ozilgen, 1988
] to describe several fermentation experiments. It assumes logistic
growth and inhibition of accumulated products.
4.1. HISTORICAL PERSPECTIVE 81
Integration of Biochemical Pathway Information and pH
The Shinto model [
Shinto et al., 2007
] proposes the first mechanistic view on the
ABE fermentation by considering the underlying biochemical pathways. Kin-
etics are Michaelis-Menten type, the CoA transferase reaction from butyrate
or acetate to the corresponding CoAs is a random bi-bi mechanism. It is a
batch fermentation model that does not distinguish between intracellular and
extracellular metabolites. Similarly to the other models it considers biomass
production that is inhibited by butanol. Here biomass production is propor-
tional to acetyl-CoA levels. [
Junne, 2010
] also proposes a dynamic model for
solvent formation. This model includes pH mediated dissociation of compounds.
Enzyme concentrations are modelled as sigmoidal function accordingly to the
available transcript expression levels during batch fermentation. Additional
transport terms for acids are implemented. Finally, models were developed that
assume the translation of enzymes is pH-dependent, these models explain the
pH-shift in the continuous chemostat experiment as proposed by the COSMIC
SOP [Haus et al., 2011,Millat et al., 2013a].
Integration of Time Series of Transcriptome Data in Kinetic Models
The integration of transcriptome data in a time series format and com-
bine it with metabolome was introduced by [
G¨otz and Reuss, 2009
], this ap-
proach seems unique in literature. A stochastic model is proposed by another
group that involves enzyme concentrations that are integrated as time profiles
[
Liebermeister and Klipp, 2006
], integration of metabolome and transcriptome
regulation in a flux balance model is described elsewhere [
Covert et al., 2001
].
A simpler approach of comparing two different concentrations is reported for a
lactate dehydrogenase model, where ratios of different isozymes are incorporated
[
Downer et al., 2006
]. In E. coli regulatory networks are inferred from mRNA
data and integrated to the mass-balance as an ODE system [
Carrera et al., 2009
].
82 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
4.2 Derivation of the Dynamical Model
The creation of a model for an organism offers many tasks. Clearly, the manifold
connections within a biological network, the interactions within one Omic and the
interaction between Omics, they all represent a tremendous amount of data to
be mapped into a model. Such model ideally covers the in vitro behaviour under
the chosen environmental conditions. However, calculation and identification
of all factors remains impossible, not only because functions of proteins remain
obscure (3.5), the amount of available data does usually not suffice to predict
a unique parameter set. Consequently, it is not the task to integrate all known
interactions, but to find a minimal model to understand the function of the
organism [Durot et al., 2009].
This chapter proposes such a model for butanol production in C. acetobutylicum
and a formalism for the implementation of time series of transcript level data.
Model Structure
Within the cell, the biochemical network of butanol synthesis is considered as
already presented (figure 2.1). Reduction of the pathways to the branching points
yields the network shown in figure 4.1. Arrows in this pathway indicate the
reaction direction as considered in the model. Touching arrows indicate the
concomitant use of several substrates to several products. Stoichiometry is not
shown in this model representation. The CAC and CAP numbers represent the
transcripts considered as relevant for the reaction. A list of compounds and used
abbreviations is given in table 4.1.
Reduction to branching points assumes that no intermediate molecule has regu-
latory functions. Since literature suggests a regulatory-role of acetyl-phosphate
and butyryl-phosphate they remain included for monitoring purposes. Glucose
uptake and glycolysis are combined to one reaction due to the fact that meas-
urements of intermediates are difficult to access. Finally, kinetics of this model
are oriented on Gheshlagi et al. [
Gheshlaghi, 2009
] and the PhD thesis of Stefan
Junne [Junne, 2010].
A structured model approach is taken here, the cell and the reactor are two separ-
ate entities. Transport phenomena within each compartment will be furthermore
neglected. Transport between compartment boundaries is simplified by assuming
an intracellular substrate is converted into an extracellular product.
The goal of this model is to monitor the current status of the cell not simply by
some pH-dependency but by following the expression of the relevant transcripts.
Therefore, no pH description is necessary and the dissociation state of acids is
unconsidered. A model of non-autonomous differential equations is achieved that
combines three types of information: the biochemical pathways, the transcript
4.2. DERIVATION OF THE DYNAMICAL MODEL 83
level data for each enzyme involved in the pathways, and the enzyme kinetic
information from literature.
11 1
7
8
2
45
3
6
0
10
9
CAC1742
CAP0165
CAC1743
CAC3076
CAC2873
CAC2708
CAC2711
CAC2712
CAP0163
CAP0164
CAC3075 CAP0162
CAC3298
CAC3299
CAP0162
CAC3298
CAC3299
Figure 4.1: Minimal Model pathway structure for butanol synthesis by
C.acetobutylicum. The grey box is the substrate 0: glucose.
Red boxes show the solvents (8: acetone, 9: ethanol, 10: butanol).
Green boxes show the acids (2: acetate, 6: butyrate). Blue boxes intracellular
intermediates (1: acetyl-CoA, 3: aceto-acetyl-CoA, 4: butyryl-CoA,5: butyryl-
phosphate, 7: acetoacetate, 11: acetyl-phosphate).
4.2.1 Derivation of the Model
Comparison of Compartment Volumes
The conversions take place in a cell, which is considered a compartment contained
in the reactor. The cell density (
X
) is defined as dry mass of cells (
mX
) per
cell volume (
VC
). The mass of cells further corresponds to the measured biomass
concentration (cX) in the reactor volume (VR):
X=mX
VC
=cXVR
VC
(4.1)
84 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
Compound number compound CHEBI-ID Abbreviation
0 glucose 17234 Glc
1 acetyl-CoA 15351 ACoA
2 acetic acid 15366 ACE
3 acetoacetyl-CoA 15345 AACoA
4 butyryl-CoA 15517 BCoA
5 butyryl-phosphate 17260 BUP
6 butyric acid 30772 BU
7 acetoacetate 13705 AA
8 acetone 15347 ACN
9 ethanol 16236 ETOH
10 butanol 28885 BUOH
11 acetyl-phosphate 15350 ACP
Table 4.1: Compound Overview
The volume of the liquid phase (
VL
) is the difference of reactor volume and cell
volume:
VL=VR−VC=VR1−VC
VR=VR1−cX
X(4.2)
Biomass of C. acetobutylicum is not growing to high cell densities, hence we can
assume
cX
X≪
1 and thereby we neglect the cell volume compared to the reactor
volume:
VL≈VR.(4.3)
Relating Reaction Rates to Cell and Reactor
The total reaction rate (
R
) is the amount of one substance (
n
) being converted
over time (
t
). Since the model is a two-compartment model, this reaction can be
referenced to the cell (ri) or to the reactor (ro).
R=riVC=roVL(4.4)
from equation 4.3 follows
ro=riVC
VR
(4.5)
from equation 4.1 follows
ro=ricX
X
(4.6)
4.2. DERIVATION OF THE DYNAMICAL MODEL 85
The Time Law for Changes of Intracellular Concentrations
In the next step, the time law for the change of an intracellular component from
the mole balance is derived. A substance
ni
is produced and consumed by
NR
reactions. The balance for an intracellular metabolite yields:
dni
dt =
NR
X
k=1
vkRk(4.7)
vk=(−1 if niis a substrate
+1 if niis a product (4.8)
Expanding the left-hand side of equation 4.7:
dni
dt =d(ciVC)
dt (4.9)
=d(ciVRcX
X)
dt (4.10)
=dci
dt
cXVR
X
+dcX
X
dt ciVR+dVR
dt
cX
X
ci(4.11)
Now assume constant cell density and exponential growth at a rate
µ
and divide
by the reactor volume:
dci
dt
cX
X
+µci
cX
X
+1
VR
dVR
dt
cX
X
ci=
NR
X
k=1
vkro
k(4.12)
Factorise cX
X
and simplify the right-hand side according to equation 4.3
cX
Xµci+dci
dt +ci
VR
dVR
dt =
NR
X
k=1
vkro
k(4.13)
For batch and chemostat, the reactor volume remains constant
dVR
dt
= 0
and
after simplification and rearrangement one gets
cX
Xµci+dci
dt =
NR
X
k=1
vkro
k(4.14)
cX
Xµci+dci
dt =
NR
X
k=1
vk
Rk
VR
(4.15)
dci
dt = (
NR
X
k=1
vkri
k)−µci.(4.16)
86 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
The Time Law for Changes of Extracellular Concentrations
As before, a substance
no
is produced or consumed by
NR
reactions. Additionally,
it is transported out of the reactor by transport term Tsink.
dno
dt =
NR
X
k=1
vkRk−Tsink (4.17)
In continuous culture this sink is due to a pump and therefor dependent on the
pump flow rate F:
Tsink =Fco(4.18)
Converting into concentrations and dividing by
VR
, where
D
=
F
VR
is the dilution
rate:
dco
dt =
NR
X
k=1
vk
Rk
VR
−F
VR
co(4.19)
=
NR
X
k=1
vkri
k−Dco(4.20)
4.2.2 Formalism
For the automated integration of the mathematical equations and biochemical
network into the model, a formalism is required that is applicable to both com-
partments such that it is only necessary to specify whether a compound is located
in one or the other.
In a first step, a compound is given an unique number either manually or by
a reckoned online repository like from the database and ontology of Chemical
Entities of Biological Interest (CHEBI). Using CHEBI-IDs is one step forward
to sustainability of the model, they are unique and searchable through online
services, and may be parsed in SysMO-SEEK.
The second step concerns the naming of reaction, as KEGG reactions are undirec-
ted and unrelated to compounds, the use of these identifiers is not recommended
here. In order to give a reaction a direction, they are called
rs|p
, where
s
and
p
are substrate-ID and product-ID delimited by the ”
|
” character. Without
loss of generality, multiple substrates and multiple products can be introduced
rs1,s2,...,sN|p1,p2,...,pM
. Equations 4.16 and 4.20 can be generalised to equation 4.21:
dck
dt =X
j
˜
Vjrj|k−X
i
˜
Virk|i−Dk·ck(4.21)
The parameter
vk
is not necessary anymore, because the direction of the
reaction is clear from the reaction identifiers,
rj|k
is the production of substance
k
,
4.2. DERIVATION OF THE DYNAMICAL MODEL 87
reaction kinetic transcripts involved
(cac:CA ...)
r0|1glucose feed
r1|11 MMT C1742
r11|1MMT C1742
r11|2MMT C1743
r2|11 MMT C1743
r2,3|1,7bi-substrate MMT P0163,P0164
r1|3bi-substrate MMT C2873
r3|4substrate inhibition C2708,C2710,C2711
r4|5competitive product inhibition C3076
r5|4MMT C3076
r5|6MMT C3075
r6|5MMT C3075
r6,3|4,7bi-substrate MMT P0163,P0164
r7|8MMT P0165
r1|9MMT C3298,C3299,P0162
r4|10 uncompetitive product inhibition C3298,C3299,P0162
r3|7r2,3|1,7+r6,3|4,7P0163,P0164
Table 4.2: Model reaction kinetics and transcripts: The
rs|p
reaction-identifier
shows the directed conversion from substrate
s
to product
p
using the transcript
data from the respective transcript-identifiers. The kinetic model kin is specified
in the second column. MMT: Michaelis-Menten type
rk|i
is the consumption of substance
k
. However, the two process designs, batch
and continuous culture, require a generalisation of the dilution
Dk
and the two
compartment system requires a the factor ˜
V.
Dtakes three values:
Dk= 0 : batch conditions, kis extracellular
Dk=µ: batch conditions, kis an intracellular compound.
Dk=F
VR
: continuous conditions, kis either extracellular or intracellular
˜
Vjand ˜
Vialso take three values:
˜
V= 0 : There is no reaction between the compounds k, i or k, j.
˜
V=ρX
cX:kis an intracellular compound.
˜
V= 1 : kis an extracellular compound.
88 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
Kinetic Laws and Integration of Transcript Levels
The amount of enzyme present in the cells is assumed to be time-dependent by
explicitly integrating time-series data of the corresponding transcripts (table 4.2)
into the rate equation. The type integration is explained in the next section, here
it is an unspecified function f.
The reaction rate further depends on a specific kinetic (kin, table 4.2) that is a
function of substrate concentrations (
cs
), product concentrations (
cp
) and in case
of inhibitions also on any other species ci.
The enzymes maximal rate
¯
ks|p
[mole/time unit] is related to this kinetic. This
rate is the product of the maximal specific rate ks|p[mole/(time unit, g biomass
and amount of enzyme)] the biomass
cX
[g biomass] and the amount of enzyme
which is a function of transcript levels
f
. This gives the generalised reaction rate
equation from a substrate sto a product p:
rs|p=¯
ks|pkins|p(cs, cp, ci) (4.22)
=ks|p·f(transcript levels of rs|p)·cX·kins|p(cs, cp, ci) (4.23)
By this definition, the established ODE-system of non-autonomous equations
constitutes a descriptive model of the ABE-process.
4.2.3 Integration of Time-Dependent Data
Description of Growth And Glucose Consumption
Growth and glucose consumption are not modelled using mass balance, nor
are they coupled to occurring reactions, e.g. acetyl-CoA was used for growth
modelling [
Shinto et al., 2007
]. They are instead implemented as piecewise linear
interpolations of the data. It is known that linear interpolations only poorly
describe the data, however non-linear estimators or functional data analysis,
require deep knowledge on data-structure and, more importantly, a sufficient large
sampling set [
Lehmann et al., 1999
,
Gustafsson et al., 2009
]. Since replicates in
fermentation experiments are rare, application of these advanced methods is
difficult.
Therefore, it is assumed that the acetyl-CoA influx is directly proportional to the
glucose uptake of the cell from the medium. The proportionality constant is the
substrate yield
Y pGlc
, it models the fraction of glucose used for growth compared
to the glucose used for metabolite synthesis.
Growth is also considered directly proportional to the measured optical density in
the medium.
Description of Transcript Levels
Transcript levels are also implemented as piecewise linear interpolations for the
same reasoning as before. Data sparseness in the temporal dimension is usually
4.2. DERIVATION OF THE DYNAMICAL MODEL 89
more severe in that data than for growth or substrate profiles. Nevertheless, other
studies were successful in establishing a transcription model that can be used for
integration of transcript levels [Chen et al., 1999].
In order to use transcript data as a protein quantity it is necessary to assume that
transcript levels map to protein levels. This assumption requires the study of two
processes, protein translation and mRNA stability. Since C. acetobutylicum is not
rapidly growing, the ribosome quantity remains constant during duplication of
the cells and is not limiting [
Golding et al., 2005
]. Second it is necessary to scale
the data to an upper bound. Since data are present in logarithmic format, the
maximum should be scaled to zero to achieve maximal flux at least once during
the time course of the experiment.
Third, in vivo transcript stability is an unknown parameter in this model. It
is a function of the cell’s status and may also differ for each transcript, a com-
plete modelling of transcript translation and degradation was carried out earlier
[
Arnold, 2002
]. It is clear from this modelling that the interplay of translation
and degradation is not a linear function, and the transcriptome time series data
needs to be shifted in a non-linear fashion in order to map to proteome data.
However, no proteome data in such temporal dimension is available and even
if it were available the here applied methods would not change. Consequently,
transcript expression data will be used as development standard until enough
protein data is available.
Lumped reactions, e.g. the three reactions from acetoacetyl-CoA to butyryl-CoA
are calculated as average of transcript levels, since usually the corresponding genes
are organised in an operon and are expected to behave similarly.
Model Equations
The entire models equations are noted in appendix A.
90 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
4.3 Model Implementation
Various tools are available to implement ordinary differential equations for simu-
lation and parameter estimation. The integration of temporal profiles from data
for transcripts, biomass and glucose represent an additional requirement to the
software.
Only two packages were found that fulfil this requirement, one is SBTOOLBOX2
[
Schmidt and Jirstrand, 2006
], a third party toolbox for MATLAB written by
Henning Schmidt. Its computing power has been widely used in the biological
community. It offers a graphical user interface that allows the execution of com-
plex calculation tasks and it offers a library of scripts that can be freely accessed.
The evaluation of C-script by MATLAB, called MEX-compilation, also greatly
increases performance of the scripts.
The commercial SimBiology Toolbox by MATLAB can be used as well, however,
its functionality is very limited when it comes to further analyses, e.g. sensitivity
analysis.
Data Pre-Requisites
Transcript data must not have no missing values. Instead of linear interpolation,
other imputation techniques are less error prone. The Metagenealyse-webpage
uses principal component analysis (PCA) for imputation [
Daub et al., 2003
] of
missing data in time-profiles of transcript data.
The Standard Format
A model standard-format is required to automatically integrate time-series data.
It should be parsable by MATLAB and convertable to a readily calculable SB-
TOOLBOX2 model (B.3).
The standard-format is sketched in a toy-model in figure 4.2. First, the basic
structure of the SBTOOLBOX2-model serves as core model and necessary para-
meters are predefined and therefor set to 0 (rGlcIn - the glucose influx into the
organism, mue - the growth rate as determined from the change of optical density,
cX - the biomass as determined from the optical density). The piecewise linear
interpolations are calculated from matrix data. Second, the implementation of
transcript data interpolation requires that transcript-identifiers (
T
1
, T
2) can be
separated and addressed. The format of a multidimensional function
f
(
T
1
, T
2)
makes this possible.
Initially, SBML models would be suitable as sustainable formats for model deposit
and retrieval, as well as interactivity between several softwares for visualisation
and calculation. However, this newly developed data-driven model type does not
fit into a SBML-standard so far. SED-ML standards may cover this shortly. Until
then this approach has to suffice.
4.3. MODEL IMPLEMENTATION 91
********** MODEL NAME
Toy Model
********** MODEL NOTES
This model serves for illustration of the automated data implementation
procedure.
********** MODEL STATES
d/dt(x)=(rhoX/cX)*(rGlcIn - rxy)
d/dt(y)=(rhoX/cX)*rxy-mu*y
x(0)=0
y(0)=0
********** MODEL PARAMETERS
rhoX=300
kxy=0.1
K=0.1
********** MODEL VARIABLES
rGlcIn=0
cX=0
mu=0
********** MODEL REACTIONS
r0x=rGlcIn
rxy=f(T1,T2)*cX*kxy*x/(x+K)
********** MODEL FUNCTIONS
********** MODEL EVENTS
********** MODEL MATLAB FUNCTIONS
Figure 4.2: Structure of the standard model in SBToolbox2: The model is
given a name that is used through out the model analysis. MODEL STATES
are the compound under investigation given by the differential equation and
its initial concentration. MODEL PARAMETERS govern the reaction kinetics.
MODEL VARIABLES are an explicit function of time. MODEL REACTIONS
are dependent on the MODEL STATES and the MODEL VARIABLES. MODEL
FUNCTIONS allow the user to supply self-defined functions.
This toy model shows the automated data implementation method. The model
consists of two differential equations for the compounds
x
and
y
, and two reactions
r0|x
and
rx|y
, rGlcIn is the glucose influx into the organism and
rx|y
a conversion
from
x
to
y
at a maximal rate of
kx|y
with Michaelis-Menten constant
K
. This
conversion rate is happening inside a compartment of time-dependent volume
cX
and constant density
X
.
µ
is the associated compartment growth rate and
governs the dilution of the compound
x
by growth. The conversion of
x
to
y
is
governed by two different transcripts
T1
and
T2
that are combined by a function
f, e.g. the average later on.
92 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
SimBiology Toolbox
As a commercially available software, the SimBiology-toolbox by MATLAB was
tested. The implementation of the model can be undertaken by a GUI guided
approach and by a scripting approach. Since the model is large, it is recommended
to use batch processing of pre-defined equation files. A list of all involved
reactions with relevant substrates and products was created and converted into an
interpretable format for this toolbox. It turned out that data-dependent variation
of transcript level profiles was not feasible ab initio in this toolbox. A programmed
work-around in C that made it possible to implement it however for all three
time courses, transcripts, biomass and glucose variation. However, it turned out
that proprietary scripts as e.g. the sensitivity analysis in SimBiology were not
supportive for varying compartment size so far. The investigation of this software
was aborted at that point. Further softwares were tested, but the same effects as
reported by Alves et al were observed: Interoperability of softwares, interfaces
and documentation are poor for many of them and specific problems were either
difficult or even impossible to be solved [Alves et al., 2006].
4.4. EVALUATION OF EXPERIMENTS 93
4.4 Evaluation of Experiments
4.4.1 Computer, Softwares, Data, Algorithms
Computer
All simulations were carried out on a AMD Phenom (TM) II X6 1090T processor
with 3.20Ghz, 16GB of RAM.
Softwares
Simulations were run on Windows 7 SP1, 64bit. The list of all softwares is given
in table 4.3.
MATLAB 7.13.0.564 (R2011b)
Bioinformatics Toolbox 4.0
Optimization Toolbox 6.1
Parallel Computing Toolbox 6.1
SimBiology 4.0
Statistics Toolbox 7.6
Symbolic Math Toolbox 5.2
SB Toolbox 2 Development
SBPD Development
Table 4.3: Softwares
Data for Parameter Estimation and Cross-Validation
The first set of data is the batch fermentation in complex medium [
Jones et al., 2008
].
Its model is called the batch model (BM). Solvent production starts after the stop
of growth in the transitional phase after 10 h.
The second set of data was collected during continuous fermentation in phos-
phate limited medium [
Grimmler et al., 2011
]. Its model is called the continuous
model (CM). For the simulation the time frame of the data was shifted since data
recording commenced only at 110
h
after inoculation. The shift occurred after
160 h.
Algorithms for Model Simulation
BM and CM are constructed by the previously described scheme (section 4.2).
Appendix Ashows the model equations for the CM, for the BM, the same equa-
tions are valid but
D
= 0. Simulation of differential equations in SBTOOLBOX2
and SBPD follows the MATLAB integration of stiff differential equations (ode15s,
ode23s) with standard parameters (absolute tolerance of 1e-6, and relative toler-
ance of 1e-3).
94 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
Algorithms for Parameter Estimation
Parameter estimation used a particle swarm search [
Vaz and Vicente, 2007
] com-
bined with the Nelder-Mead-Simplex algorithm [
Flannery, 2007
]. The first al-
gorithm offers the possibility to find a global minimum of the optimization
functional because of its stochastic nature. For the same reason, its convergence
to the minimum is very slow and therefore enhanced by the simplex algorithm.
Both algorithms are implemented to consider box constraints of the parameters. It
was not possible to impose constraints on the states. The estimation is carried out
only for the maximal rates and the substrate yield within a box of 10
−3mM
to
10
3mM
. Michaelis-Menten constants were set to 1
mM
and inhibitory constants
were set to 1 M.
Approach for Simulation of Mutation Experiments
Experiments in which promoters were replaced by a new promoter are mapped to
the model by replacing the transcript level profiles of the original promoter with
the corresponding profiles of transcripts behind the new promoter. This profile
transfer is feasible under the assumption that the newly integrated promoter
and the changed transcript expression does not profoundly affect the overall
expression.
Knock-down and over-expression studies are simulated by assuming that the
dynamics of the transcript-data persist, but its levels are changed. This is done
by alternating the maximal conversion rate ks|p.
For deletion experiments the corresponding rate was decreased by three orders of
magnitude because low values of activities are reported even for deletion mutants.
For up-regulation experiments the double of the corresponding rate was assumed,
as [Mann and Luetke-Eversloh, 2013] indeed reported.
Approach for Disturbance Analysis
For estimation of parameter certainty a disturbance analysis is carried out. The
initial parameter-set is perturbed by 10% and re-estimated using the Nelder-Mead-
simplex algorithm. This is done several times, here
n
= 200, and the resulting
parameter-distributions after re-estimation are evaluated.
Approach for Cross-Validation
A cross-validation by using batch and continuous data is carried out. The
corresponding models, BM and CM, are re-estimated within a 20% range of
the original parameter set. Only the substrate yield
Y pGlc
was permitted to
re-adjust freely, assuming that different media compositions strongly affect glucose
consumption.
4.4. EVALUATION OF EXPERIMENTS 95
List of Goals for Validation
The following list of facts is tested for validation of the model and its simulation
results.
1.
Levels and dynamics of measured metabolomic data of wild-type cultures
[Jones et al., 2008,Grimmler et al., 2011] are predicted.
2.
Levels and dynamics of measured metabolomic data of mutant-type
cultures [
Green et al., 1996
,
Lehmann et al., 2012a
,
Lehmann et al., 2012b
,
Mann and Luetke-Eversloh, 2013] are predicted.
3.
Cross-Validation of continuous data and batch data leads to comparable
results.
4.
In batch culture in complex medium butyryl-CoA and butyryl-phosphate
show twin-peaks, acetyl-phosphate shows one peak, corresponding to acid
uptake [Zhao et al., 2005,Amador-Noguez et al., 2011].
5.
Acetyl-CoA and butyryl-CoA concentrations decrease during the shift in
continuous culture [Grupe and Gottschalk, 1992].
6.
In batch culture the amounts of acetoacetyl-CoA, 3-hydroxybutyryl-
CoA, crotonyl-CoA are less than 21
µM
, 11
µM
, 10
µM
, respectively
[Boynton et al., 1994].
7.
The following metabolites are higher concentrated in the mid-exponential
phase than in the solventogenic phase: acetyl-phosphate, acetyl-CoA,
butyryl-CoA, 3-hydroxybutyryl-CoA. Acetyl-CoA and butyryl-CoA pro-
files are similar [Amador-Noguez et al., 2011].
8.
Acetate concentrations and acetone production are correlated but not butyr-
ate concentrations, the ATP balance is favoured via acetate phosphorylation
[Desai et al., 1999,Lehmann et al., 2012b].
9.
The CoA-transferase preferably acts on acetate. Acetate kinase is favoured
in the reverse direction. Phosphotransbutyrylase and thiolase have the
highest activity [Vasconcelos et al., 1994].
10.
Activity of phosphotransbutyrylase is seen earlier than butyrate kinase in
batch fermentations. Thiolase and
β
-hydroxybutyryl-CoA dehydrogenase
peak in mid-exponential phase [Hartmanis and Gatenbeck, 1984].
11.
The pools of acetyl-CoA and acetate are highly interchangeable because of
the rapid reversibility of phosphotransacetylase [
Amador-Noguez et al., 2011
].
12.
Parallel activity of solvent and acid pathways in one organism are unfavour-
able and indicate a mixed population [Clarke et al., 1988].
96 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
4.4.2 Parameter Estimations and Validation
Results of the Batch Model
0 20 40 60 80
0
2
4
x 104
time [h]
concentration [mM]
ACoA
AACoA
BCoA
BUP
ACP
AA
0 20 40 60 80
0
100
BM
ACE
BU
0 20 40 60 80
0
100
ACN
EtOH
BuOH
0 50 100 150
0
5
10
time [h]
0 50 100 150
0
50
BM Validation
0 50 100 150
0
50
Figure 4.3: Parameter estimation of the batch culture metabolome data using the
BM and cross-validation to the CM and continuous culture data. Model results
(left) are split into acids(above), solvents (center) and intracellular metabolites
(below). The validation (right) is split accordingly. Dotted values represent data,
lines represent model results.
Metabolite profiles are shown in figure 4.3 and reaction profiles in figure 4.4.
The simulation of concentrations of acetic acid and butyric acid in the BM follow
the dynamics of the data. Levels of simulated butyric acid are too low compared to
measurement data and the peak at 18h cannot be reached. The simulated solvent
levels of acetone and ethanol correspond to measurements, however the butanol
production starts too early and does not reach the maximal level. The intracellular
metabolites acetyl-CoA, acetoacetate and butyryl-CoA reach an unrealistic high
level in the simulation. Acetoacetyl-CoA shows a short spike to 1
M
around 8
h
.
The simulated concentrations of acetyl-phosphate and butyryl-phosphate reach a
plateau around 0.5Mand 0.25 Mrespectively during the exponential phase.
The sum of the CoA-transferase reactions
r3|7
is highest at 18
h
, the onset of
solventogenesis. The major contribution to this reaction is given by
r63|47
not
by
r23|17
. Activity of butanol dehydrogenase
r4|10
shows a similar profile to
acetoacetate formation, however there is an earlier peak. Acetone formation is
bimodal, one peak at 18h, the second between 30
h
and 40
h
. Ethanol formation
is very low overall.
r4|5
and
r5|4
are the fastest reaction, they have an equal and again bimodal
behaviour: one peak at 18
h
and a second 4
h
later.
r1|3
has a peak at the
same time, the further processing via
r3|4
is very low, with a maximum earlier
4.4. EVALUATION OF EXPERIMENTS 97
0 20 40 60 80
0
50
100
150
BM
reaction rate [mM/h]
0 20 40 60 80
0
10
20
time [h]
0 50 100 150
0
20
40
BM - Validation
r1|11
r11|1
r11|2
r2|11
r1|3
r3|4
r4|5
r5|4
r5|6
r6|5
0 50 100 150
0
1
2
3
time [h]
r23|17
r63|47
r7|8
r1|9
r4|10
r3|7
Figure 4.4: Parameter estimation of the batch culture data metabolome using the
BM and cross-validation to the CM and continuous culture data. Model results
(left) are split into reactions known to be active during acidogenesis (above) and
during solventogenesis (below). The validation (right) is split accordingly.
around 8
h
. Both reverse kinase reactions
r6|5
and
r2|11
are not detectable.
The forward direction of the kinases is similar during acidogenic conditions,
only the acetate kinase reaction reappears during solventogenic conditions. The
phosphotransacetylase reaction performs equally well in both directions (
r1|11
and
r11|1).
Results of the Continuous Model
Metabolite profiles are shown in figure 4.5 and reaction profiles in figure 4.6.
Estimation of acids and solvents concentrations in the CM follow the dynamics of
the pH-shift. A sharp decrease to zero and almost zero is calculated for butyrate
and acetate concentrations, respectively. Accordingly, the solvents increase to the
levels given by the measurement data. Acetoacetate concentration is highest with
a peak of more than 2
mM
when butyrate uptake stops. Its level continue to be
high around 1
mM
during solventogenesis. Butyryl-CoA and butyryl-phosphate
concentrations show an elevation during the shift to 1
mM
and to 0
.
25
mM
respectively. The acetyl-CoA concentrations is shortly decreased during the shift
and regains its previous value of around 0.25 mM.
As in the batch model, the acetoacetate production
r3|7
is dependent only on
r63|47
and not on
r23|17
. The decarboxylation
r7|8
has the same velocity as
r3|7
.
Production of alcohols occurs in the expected split ratio 1:3 approximately.
As before,
r4|5
,
r5|4
are the fastest reactions, they are decreasing during the shift
98 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
0 50 100 150
0
1
2
time [h]
concentration [mM]
ACoA
AACoA
BCoA
BUP
ACP
AA
0 50 100 150
0
50
CM
ACE
BU
0 50 100 150
0
50
ACN
EtOH
BuOH
0 20 40 60 80
0
5
x 104
time [h]
0 20 40 60 80
0
100
CM -Validation
0 20 40 60 80
0
100
Figure 4.5: Parameter estimation of the continuous culture metabolome data using
the CM and validation using the BM. Model results (left) are split into acids(above),
solvents (center) and intracellular metabolites (below). The validation (right) is
split accordingly. Dotted values represent data, lines represent model results.
0 50 100 150
0
20
40
60
80
CM
0 50 100 150
1
2
3
time [h]
reaction rate [mM/h]
0 50 100 150
0
100
200 CM Validation
r1|11
r11|1
r11|2
r2|11
r1|3
r3|4
r4|5
r5|4
r5|6
r6|5
0 50 100 150
0
10
20
time [h]
r23|17
r63|47
r7|8
r1|9
r4|10
r3|7
Figure 4.6: Parameter estimation of the continuous culture metabolome data using
the CM and validation using the BM. Model results (left) are split into reactions
known to be active during acidogenesis (above) and during solventogenesis (below).
The validation (right) is split accordingly.
4.4. EVALUATION OF EXPERIMENTS 99
and increase afterwards. This behaviour is seen also for
r1|3
,
r11|1
,
r1|11
, the others
decline after the shift. r5|6is stronger than r6|5.
Validation of the Batch Model
Metabolite profiles are shown in figure 4.3 and reaction profiles in figure 4.4.
The simulation results of the BM are comparable to the CM calculations: Acid
and solvent concentrations are met, with the exception of acetone. The dynamics
of solvent formation are also met. For the intracellular metabolites, again a huge
acetoacetate signal is seen, however, this time there is no accumulation of butyryl-
CoA, but still a twin-peak of 0
.
3
mM
height. The acetoacetyl-CoA concentration
is step-wise increasing, which is in contrast to the CM. The reason for this is the
CoA-transferase reaction
r3|7
, that has reduced activity during solventogenesis
compared to the CM. The other solventogenic reactions look similar. During
acidogenesis
r4|5
and
r5|4
are still the fastest reactions. The other reaction rates
are comparable to the rates in the CM.
Validation of the Continuous Model
Metabolite profiles are shown in figure 4.5 and reaction profiles in figure 4.6.
The dynamics of solvent production are met by the validation model, however
levels for butanol are underestimated and for acetone and ethanol overestimated.
The dynamics of acid concentrations are well met for acetate but levels are
largely underestimated. The dynamics of butyrate concentrations are not met
and levels are underestimated. The validation model shows the same huge spikes
of intracellular metabolites as the BM. The other intracellular concentrations
and dynamics are similar to the BM, the plateau of butyryl-phosphate is more
elongated until 40
h
. The reactions involved in solventogenesis have similar
dynamics and levels as the CM. The twin-peaked reactions
r4|5
and
r5|4
are
appearing as a single peak with its maximum higher than in the CM.
Parameter Sets and Disturbance Analysis
Considering the parameters (table 4.4), it is obvious that the conversaion rate
for acetyl-CoA dehydrogenation to acetaldehyde and ethanol (
k1|9
) is two orders
of magnitude lower than for butyryl-CoA (
k4|10
), still there is accumulation of
butyryl-CoA in the CM.
k1|3
,
k3|4
,
k4|5
,
k5|4
have the highest specific rates and the
reverse rate
k4|5
is stronger than the forward rate in every model. This is similar
for the acetate production branch:
k11|1
is higher than
k1|11
. Not surprisingly
from the simulation curves
k63|47
is large and the values of
k23|17
,
k6|5
and
k11|2
are close to zero. The major difference between the two parameter sets of CM
and BM is the activity of acetate kinase, the reverse direction
k6|5
is active in the
continuous culture but not in the batch culture. Not surprisingly, many of the
100 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
Table 4.4: Parameter-Estimations of batch culture data and continuous culture
data with the corresponding cross-validation of CM and BM and vice-versa within
a 20% margin of the parameters
parameter CM cross-validation BM cross-validation
parameters of CM to BM parameters of BM to CM
Y pGlc 6.07E-01 1.80E+00 1.42E+00 4.80E-01
k1|11 3.30E+01 3.95E+01 2.72E+01 2.17E+01
k11|13.29E+02 2.64E+02 7.36E+01 8.84E+01
k11|28.69E+01 1.04E+02 6.27E+01 5.01E+01
k2|11 1.00E-04 8.03E-05 1.74E-05 2.09E-05
k23|17 1.00E-04 1.20E-04 1.64E-05 1.97E-05
k1|32.02E+02 2.55E+02 1.91E+02 2.29E+02
k3|49.61E+02 8.31E+02 8.28E+02 6.62E+02
k4|51.31E+02 1.05E+02 1.65E+02 1.97E+02
k5|43.72E+02 4.47E+02 6.19E+02 4.95E+02
k5|69.78E+01 7.83E+01 8.45E+01 1.01E+02
k6|51.49E+00 1.19E+00 1.26E-05 1.49E-05
k63|47 1.97E+02 2.36E+02 1.94E+02 1.93E+02
k7|86.08E+00 4.86E+00 4.17E+00 5.00E+00
k1|93.89E+00 3.11E+00 1.21E+00 1.46E+00
k4|10 2.94E+02 3.53E+02 1.03E+03 8.23E+02
estimated parameters are highly uncertain, as a disturbance analysis indicates
(figure 4.7): The major part of parameters shows an approximately 10% variance
after the disturbance, the parameters
k63|47, k3|4
have a lower variance, and the
parameters
k1|9, k6|5, Y pglu
the lowest variance. Asymmetries in the parameter
sets’ variances indicate that the objective function is asymmetrically shaped, e.g.
k3|4.
Validation by Mutant Experiments
Promoter experiments are shown in figure 4.8, deletion studies are given in figure
4.9.
4.4. EVALUATION OF EXPERIMENTS 101
0.8 0.9 1 1.1 1.2 1.3
(0.507 ± 0.000312) Ypglu
(31.8 ± 3.91) k1|11
(334 ± 40.1) k11|1
(83.9 ± 10.2) k11|2
(9.36e−005 ± 1.26e−005) k2|11
(9.37e−005 ± 1.22e−005) k23|17
(206 ± 24.1) k1|3
(815 ± 55.8) k3|4
(138 ± 15.8) k4|5
(397 ± 45.6) k5|4
(92 ± 11.5) k5|6
(1.82 ± 0.0632) k6|5
(217 ± 12.7) k63|47
(6.36 ± 0.725) k7|8
(3.59 ± 0.198) k1|9
(280 ± 33.6) k4|10
Figure 4.7: Boxplot of the parameter uncertainties in the CM after a 10%
disturbance of the initial parameter set. Parameters are scaled to their mean
value.
102 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
0
100
200
0
100
200
conc. [mM]
0
100
200
0 20 40 60
0
100
200
time [h]
0 20 40 60
time [h]
ACE
BU
ACN
EtOH
BuOH
A
B
C
D
Figure 4.8: Model results for published batch experiments of mutant cultures
whose transcripts are expressed behind a different promoter:
(A): wild type results
(B): silenced ctfB1 [
Tummala et al., 2003b
]. The authors report low butanol and
acetone titers and high butyrate titers. Here, butanol titers are unchanged and
indeed butyrate titers are elevated.
(C): alcohol dehydrogenase under the phosphotransbutyrylase promoter
[
Sillers et al., 2009
]. The authors report enhanced butanol and ethanol yields,
which is covered by the model. However, acetate does not accumulate as expected.
(D): both modification from (B) and (C) [Sillers et al., 2009]. Author report the
highest solvent yields in this mutant and best re-uptake of butyrate. Here butyrate
is taken up better than in (B) but still elevated levels are seen. Solvent yields are
indeed a little higher.
4.4. EVALUATION OF EXPERIMENTS 103
0
100
200
0
100
200
0
100
200
conc. [mM]
0
100
200
0
100
200
0 20 40 60
0
100
200
time [h]
0 20 40 60
time [h]
ACE
BU
ACN
EtOH
BuOH
A
B
C
D
E
F
Figure 4.9: Model results for several published batch experiments with deletion
mutants and optimised enzyme activities:
(A): phosphotransacetylase deletion [
Lehmann et al., 2012a
]. Authors report no
drastic change is provoked through, here ethanol yields are increased.
(B): CoA-transferase and acetoacetate decarboxylase deletion
[
Lehmann et al., 2012a
]. The authors report high amounts of acetate in
the fermentation broth, this cannot be recovered. Here, butyrate is accumulating
and the solvent production is similar to (A).
(C): CoA-transferase, acetoacetate decarboxylase and phosphotransacetylase
deletion [
Lehmann et al., 2012a
]. The authors report drastically decreased
acetate concentrations, this cannot be recovered. Here, solvent production is
similar to (A).
(D): phosphotransbutyrylase deletion [
Lehmann et al., 2012b
]. The authors
report elevated ethanol and butanol titers, this is covered by the model.
(E): enhanced thiolase activity [
Mann and Luetke-Eversloh, 2013
]. The authors
report elevated ethanol and butanol titers by 50% to 19% respectively. Butanol
titers are unchanged here, but ethanol elevated.
(F): butyrate kinase knock-down [
Green et al., 1996
]. The authors report reduced
and delayed butyrate formation and increased butanol production. This is
recovered by the model.
104 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
4.4.3 Validation-Results
1.
Levels and dynamics of the measured data are very well covered by the CM,
they are well covered by the BM. This expectation is met.
2.
Mutant culture experiments can be partly mapped without further model
adjustment. This expectation is partly met.
3.
Cross-validation is able to cover the dynamics but not the levels of the
metabolite data. This expectation is partly met.
4.
Plateaus not peaks of butyryl-phosphate and acetyl-phosphate are observed
in the BM. Since these plateaus coincide with the overflow peak of acetyl-
CoA, there may be saturation of the reactions. This expectation may be
met.
5.
Effectively, acetyl-CoA and butyryl-CoA concentrations do decrease during
the shift in the CM. This expectation is met.
6.
The amount of acetoacetyl-CoA in the BM is larger than expected by ap-
proximately one magnitude. On the one hand, this could be explained to the
overflow of acetyl-CoA, on the other hand, overnight incubation of cell pel-
lets may have disrupted this intermediate [
Grupe and Gottschalk, 1992
], as
intracellular metabolites usually have a high turn-over and rapid disruption
[Schaub, 2005]. This expectation may be met.
7.
The acetyl-CoA pool in the BM is very similar to the butyryl-CoA pool,
although they are one hour apart. In the CM, intracellular metabolites are
approximately equal concentrated in both phases. This expectation is met.
8.
Butyrate concentrations are uniquely correlated to acetone production in
both models, and acetate concentrations are only diverted via the kinase
pathway. In order to overcome this drastic difference it is necessary to
incorporate ATP generation and consumption into the model. This would
allow to valorise the more profitable acetyl-phosphate generation on expense
of the butyryl-phosphate production. This expectation is not met.
9.
The reverse phosphotransferase reaction is favoured in both models, however
the reverse kinase reaction can not be seen. This expectation is not met.
10.
A high peak of thiolase activity and a smaller peak for 3-hydroxybutyryryl-
CoA dehydrogenase can be effectively seen in the BM. The kinases and
phosphotransferases however appear in parallel. This expectation is partly
met.
4.4. EVALUATION OF EXPERIMENTS 105
11.
High interchangeability of pools is given for the butyrate production from
butyryl-CoA. This may be due as well to the symmetric structure of the
model and be overcome integrating ATP. This expectation is not met.
12.
Parallel occurrences of solventogenic and acidogenic pathways can be seen
in both, the CM and the BM. This expectation is met when a mixed culture
were present.
106 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
4.5 Sensitivity Analysis
The parametrised dynamic model will be used to identify bottle necks from
which optimisation approaches are derived. Sensitivity analysis (SA) has aided
in the identification of key factors in several biological and chemical models
[
Saltelli et al., 2000
,
Cho et al., 2003
,
Bentele et al., 2004
,
Lebedeva et al., 2012
,
Pagel et al., 2013
]. The local SA (LSA) focusses on a single point in parameter
space, the global SA (GSA) on parameter regions. A complete introduction to
both approaches is given elsewhere [Cacuci, 2003].
Local methods provide a high level of detail with less extensive calculations,
whereas global methods are best suited to handle highly variable parameters at the
cost of higher calculation demand [
Rabitz et al., 1983
]. The possibility of LSA to
only vary one parameter at a time neglects the opportunity to study the non-linear
effects of the different parameters amongst each other [
van Riel, 2006
], through this
the LSA is only able to capture few characteristics of the system [
Zi et al., 2005
].
It is indeed reported that solutions of the LSA appear as subsets of solutions
of the GSA [
Lebedeva et al., 2012
], which usually varies several parameters at
the same time. GSA is used for model simplification and it aids in parameter
estimation, it is better suited to problems in which uncertainties are in the order
of magnitudes and a strong non-linearity exists. Nevertheless, high numbers of
input variables and parameters are difficult to treat and a focus on a specific
model problem is recommended [Saltelli et al., 2000].
4.5.1 Local Sensitivity Analysis (LSA)
Given is a system of ordinary differential equations, with states
x
and parameters
p
dx
dt =fx,p, t,x(0) = x0(4.24)
Local sensitivities indices are calculated by evaluating the partial derivatives
snq
defined and normalised as follows:
s∗
nq =pq
xn
·δxn(t)
δpq
, snq(0) = 1 (4.25)
A state
xn
is called sensitive when its sensitivity is large, it is insensitive when
the sensitivity is close to zero. Since it is assumed, that all states are sensitive at
t= 0, insensitive states will decline towards zero over time.
Implementation
Several mathematical methods exist to calculate these equations [
Rabitz et al., 1983
].
The direct differential method is implemented here, because its calculation is facil-
itated through the Symbolic Toolbox of MATLAB and SBTOOLBOX2. A script
4.5. SENSITIVITY ANALYSIS 107
was programmed to integrate all sensitivity index equations into a SBTOOLBOX2
model (B.3.4).
LSA of the Continuous Model
The pathways are given in figure 4.1. A computation of the local sensitivity indices
of the CM shows that the sensitivity of acetate high for the phosphotransacetylase
forward-reaction (
r1|11
) and the acetate kinase (
r11|2
) during acidogenesis but
not during solventogenesis. No other metabolite is sensitive to these two reac-
tions (figure 4.10). Similarly, no metabolite is sensitive to the CoA-transferases
(
r63|47, r23|47
), the thiolase (
r1|3
) as shown in figure 4.11, or the lumped reaction
r3|4
and the transbutyrylase (
r4|5, r5|4
) as shown in figure 4.12. Butyrate and
butanol are sensitive to the reverse reaction of the butyrate kinase (
r6|5
) during
solventogenesis as shown in figure 4.13 and ethanol is highly sensitive to the
dehydrogenases (r1|9) as shown in figure 4.14.
This finding suggests that on the one hand that only
k6|5
and
k1|9
are sensitive
and an effect from variation of these parameters may be expected in the close
proximity. On the other hand, this finding corresponds to the uncertainty analysis
carried out earlier (figure 4.7) and suggests that these two parameters are the
most certain ones.
−2
0
2
k1|11
−2
0
2
k11|1
0 20 40 60 80
−2
0
2
k11|2
time [h]
Sensitivity
ACE
BU
0 20 40 60 80
time [h]
ACN
EtOH
BuOH
Figure 4.10: LSA of the upper branch of acid formation. Sensitivities of acids
(left) and solvents (right) are shown as function of
k1|11
(above),
k11|1
(middle)
and k11|2(below).
108 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
−2
0
2
k23|17
0 20 40 60 80
−2
0
2
k1|3
time [h]
Sensitivity
ACE
BU
0 20 40 60 80
time [h]
ACN
EtOH
BuOH
−2
0
2
k63|47
Figure 4.11: LSA of the CoA-transferase and the thiolase. Sensitivities of acids
(left) and solvents (right) are shown as function of
k63|47
(above),
k23|17
(middle)
and k1|3(below).
−2
0
2
k3|4
−2
0
2
k4|5
0 20 40 60 80
−2
0
2
k5|4
time [h]
Sensitivity
ACE
BU
0 20 40 60 80
time [h]
ACN
EtOH
BuOH
Figure 4.12: LSA of the lower branch of acid formation.Sensitivities of acids (left)
and solvents (right) are shown as function of
k3|4
(above),
k4|5
(middle) and
k5|4
(below).
4.5. SENSITIVITY ANALYSIS 109
−2
0
2
k5|6
0 20 40 60 80
−2
0
2
k6|5
time [h]
Sensitivity
0 20 40 60 80
time [h]
ACN
EtOH
BuOH
ACE
BU
Figure 4.13: LSA of butyrate kinase. Sensitivities of acids (left) and solvents
(right) are shown as function of k5|6(above) and k6|5(below).
−2
0
2
k7|8
−2
0
2
k1|9
0 20 40 60 80
−2
0
2
k4|10
time [h]
Sensitivity
ACE
BU
0 20 40 60 80
time [h]
ACN
EtOH
BuOH
Figure 4.14: LSA of dehydrogenases. Sensitivities of acids (left) and solvents
(right) are shown as function of k7|8(above), k1|9(middle) and k4|10 (below).
110 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
4.5.2 Global Sensitivity Analysis (GSA)
The parameter-set is uncertain as was shown earlier, also the LSA indicates that
the sensitivity of all states to all parameters is small. In the close proximity of
this parameter set there is no possibility to enhance the productivity of the cell
or to alleviate a possible bottle-neck. In order to estimate the global behaviour
of the model, a huger parameter region is now considered. Such evaluation is
possible by several methods that are all suited for GSA.
Methods and Implementation
Published methods for global sensitivity analysis are:
•Fourier Amplitude Sensitivity Test (FAST) [Saltelli et al., 1999]
•Partial Rank Correlation Coefficient (PRCC) [Bentele et al., 2004]
•Sobols Method [Sobol, 2001].
Applications of these three methods can be found in [
Zheng and Rundell, 2006
,
Marino et al., 2008
]. PRCC answers the question how much a model output is
dependent on the parameter, while FAST indicates which parameter uncertainty
has the highest influence on the model variance. Sobols method and FAST are
comparable. In the scope of discrete transcriptomic data that induces steps into
the model, FAST seems the best suited algorithm because it is suited also for non-
monotonic systems. PRCC is not accurate for such systems [
Marino et al., 2008
].
The FAST Algorithm
This section is summarising [Saltelli et al., 1999].
Sensitivities by FAST represent fractions of the variance
Dp
caused by varying a
parameter to the overall variance D.
The variance is the second moment of a summary statistic over the
Np
-dimensional
parameter space
KNp
= (
p|
0
≤pq≤
1;
q
= 1
, ..., Np
). More generally, the
r
th
moment of the ODE-system fis given by:
< x(r)>=ZKn
fr(p)P(p)dp(4.26)
where
P
is a probability distribution function over the parameters. The first
step is the calculation of such a statistic, by exploring the parameter space
KNp
such that a filling curve
pq
(
s
) =
Gq
(
sin ωqs
) with
−∞ ≤ s≤ ∞
comes close
to any point within.
Gq
is a transformation and the
ωq
are properly selected
frequencies.
KNp
is filled entirely if only the frequencies are incommensurate:
4.5. SENSITIVITY ANALYSIS 111
These frequencies cannot be obtained by a linear integer combination of the other
frequencies:
Np
X
q=1
rqωq6= 0,−∞ < rq<+∞(4.27)
Having distributed the parameters identically and uniformly, the integral for the
r
th moment (equation 4.26) is simplified and evaluated along the filling curves
pq
:
¯x(r)= lim
T→∞
1
2TZT
−T
fr(p(s))ds (4.28)
However incommensurate frequencies cannot be achieved due to numeric precision.
Hence, there is a
T
for which
f
(
s
) =
f
(
s
+
T
) and it was shown that if
ωq
are
positive integers, T= 2π.
The total variance Dof the model is therefore given by
D= ¯x(2) −¯
x(1)2=1
2πZπ
−π
f2(s)ds −1
2πZπ
−π
f(s)ds2(4.29)
Expanding
f
(
s
) in a Fourier series over the domain of integer frequencies
j
with
its spectrum
f(s) = P+∞
j=−∞ Ajcos(jsp) + Bjsin(js) (4.30)
Aj=1
2πRπ
−πf(sp) cos(js)ds (4.31)
Bj=1
2πRπ
−πf(sp) sin(js)ds (4.32)
Λj=A2
j+B2
j(4.33)
f
(
s
) is real valued so that the variance attributed to the fundamental frequency
ωqand its higher harmonics hωqcan be written as
Dq= 2
+∞
X
h=1
Λhωq(4.34)
The ratio Dp
Dis the estimate of the main effect of pqon x.
Implementation
FAST, PRCC and Sobols method are readily implemented in SBTOOLBOX2.
However, all three methods do not allow a temporal resolution of the sensitivities.
The corresponding scripts were adopted in order to allow the calculation of sensit-
ivity indices over time (B.3.4).
The algorithm was set to calculate the sensitivities in the parameter cube centred
around the original parameter set with an upper-boundary two-fold larger and a
lower boundary only half of the original parameter set. Sensitivities are calculated
over time-intervals, here the first interval was chosen to last until 40
h
and then
hourly steps until 80
h
were calculated until then the last interval took until 100h.
112 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
GSA of the Continuous Model
Acetate concentrations are weakly sensitive to the transacetylase (
k1|11, k11|1
)
and acetate kinase (k11|2) but to no other metabolite (figure 4.15).
Increasing sensitivity indices of butyrate and acetone towards
k63|47
during solvent-
ogenesis indicate that only butyrate-uptake but not acetate-uptake can be influ-
enced via the CoA-transferase pathway. Ethanol is weakly sensitive to thiolase
(k1|3) (figure 4.16).
Acetone sensitivity is decreasing for
k3|4
from high values in acidogenesis to small
values in solventogenesis, the profile of acetate is increasing during solventogenesis
to medium values. Sensitivity to the transbutyrylase (
k4|5, k5|4
) is decreasing for
butyrate from medium values in acidogenesis to small values in solventogenesis, it
is increasing for acetone to medium values in solventogenesis, it is constant for
butanol again at low values (figure 4.17).
The forward reaction of butyrate kinase (
k5|6
) has a weak influence on butyrate
during acidogenesis and on butanol for the whole fermentation. Again, sensitivity
of acetone is increasing for the forward-reaction of the kinase but not the reverse
direction during solventogenesis (figure 4.18).
Sensitivity of ethanol to
k1|9
is largest, suggesting that ethanol yields are easily
influenced while the sensitivity of butanol towards
k4|10
is production rate is only
average throughout the entire fermentation. Sensitivity of acetate peaks weakly
for the same parameter at 60h(figure 4.19).
0
0.5
1
d(ACE)/d(k11|1)
d(BU)/d(k11|1)
d(ACN)/d(k11|1)
d(EtOH)/d(k11|1)
d(BuOH)/d(k11|1)
0
0.5
1
d(ACE)/d(k11|2)
d(BU)/d(k11|2)
d(ACN)/d(k11|2)
d(EtOH)/d(k11|2)
d(BuOH)/d(k11|2)
0 20 40 60 80
0
0.5
1
time [h]
Sensitivity
d(ACE)/d(k1|11)
d(BU)/d(k1|11)
0 20 40 60 80
time [h]
d(ACN)/d(k1|11)
d(EtOH)/d(k1|11)
d(BuOH)/d(k1|11)
Figure 4.15: GSA of the upper branch of acid formation.
4.5. SENSITIVITY ANALYSIS 113
0 20 40 60 80
0
0.5
1
time [h]
d(ACE)/d(k1|3)
d(BU)/d(k1|3)
0 20 40 60 80
time [h]
d(ACN)/d(k1|3)
d(EtOH)/d(k1|3)
d(BuOH)/d(k1|3)
0
0.5
1
Sensitivity
d(ACE)/d(k23|17)
d(BU)/d(k23|17)
d(ACN)/d(k23|17)
d(EtOH)/d(k23|17)
d(BuOH)/d(k23|17)
0
0.5
1
d(ACE)/d(k63|47)
d(BU)/d(k63|47)
d(ACN)/d(k63|47)
d(EtOH)/d(k63|47)
d(BuOH)/d(k63|47)
Figure 4.16: GSA of the CoA-transferase and thiolase.
0
0.5
1
d(ACE)/d(k3|4)
d(BU)/d(k3|4)
d(ACN)/d(k3|4)
d(EtOH)/d(k3|4)
d(BuOH)/d(k3|4)
0 20 40 60 80
d(ACN)/d(k5|4)
d(EtOH)/d(k5|4)
d(BuOH)/d(k5|4)
0
0.5
1
Sensitivity
d(ACE)/d(k4|5)
d(BU)/d(k4|5)
0 20 40 60 80
0
0.5
1
d(ACE)/d(k5|4)
d(BU)/d(k5|4)
d(ACN)/d(k4|5)
d(EtOH)/d(k4|5)
d(BuOH)/d(k4|5)
Figure 4.17: GSA of the lower branch of acid formation.
0
0.5
1
Sensitivity
d(ACE)/d(k5|6)
d(BU)/d(k5|6)
d(ACN)/d(k5|6)
d(EtOH)/d(k5|6)
d(BuOH)/d(k5|6)
0 20 40 60 80
0
0.5
1
time [h]
d(ACE)/d(k6|5)
d(BU)/d(k6|5)
0 20 40 60 80
time [h]
d(ACN)/d(k6|5)
d(EtOH)/d(k6|5)
d(BuOH)/d(k6|5)
Figure 4.18: GSA of the butyrate kinase.
114 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
0
0.5
1
d(ACE)/d(k7|8)
d(BU)/d(k7|8)
d(ACN)/d(k7|8)
d(EtOH)/d(k7|8)
d(BuOH)/d(k7|8)
d(ACN)/d(k1|9)
d(EtOH)/d(k1|9)
d(BuOH)/d(k1|9)
0
0.5
1
d(ACE)/d(k1|9)
d(BU)/d(k1|9)
0 20 40 60 80
0
0.5
1
time [h]
Sensitivity
d(ACE)/d(k4|10)
d(BU)/d(k4|10)
0 20 40 60 80
time [h]
d(ACN)/d(k4|10)
d(EtOH)/d(k4|10)
d(BuOH)/d(k4|10)
Figure 4.19: GSA of the dehydrogenases.
4.5. SENSITIVITY ANALYSIS 115
4.5.3 Summary
Not surprisingly, LSA and GSA obtained different results from the CM: While the
LSA focusses on a single point in parameter space, the global analysis summarises
the behaviour of the model in a parameter cube.
In both simulations ethanol is highly sensitive to
k1|9
indicating that it is easy
to shift the strain to an ethanol producer, which is proven by experiments
[
Lehmann and Luetke-Eversloh, 2011
,
Lehmann et al., 2012a
]. Acetate concen-
trations are not sensitive to the activity of CoA-transferases. While the local
sensitivity points out the sensitivity of the reverse butyrate kinase reaction (
k5|6
)
alone, the global sensitivity analysis shows that cycling of butyrate through the
CoA-transferase and the transbutyrylase/kinase pathway influences the model.
116 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
4.6 Final Conclusions
In this chapter a dynamical model was presented that integrated transcriptome
data with the biochemical reaction network of solvent production. Two dif-
ferent experiments, a continuous culture experiment under phosphate limiting
conditions and a batch culture experiment in complex medium were used for para-
meter estimation and cross-validation. Biological data and mutation experiments
were qualitatively mapped to the model and then two sensitivity analyses, the
local and the global sensitivity analysis, were conducted to identify bottle necks
and engineering strategies. High expectations are imposed on in silico models
[
Lee et al., 2008b
]. Universal applicability and predictive capability are only two
of them.
Universal Applicability of the Dynamic Model
The proceeding of model development and evaluation in this thesis was undertaken
semi-automatically. Download of relevant information from KEGG and the cre-
ation of a standard format for SBTOOLBOX2 allows the automated integration of
any reaction network with any data, not only transcriptome but also proteome data.
The selection of desired sub-networks is manual work. As soon as published curated
reactomes become available as download [
Kumar et al., 2012
,
Agren et al., 2013
]
entire automation is possible with this script or the SBML interface of SBTOOL-
BOX2.
The model was able to cover two entirely different experimental proceedings
(batch culture versus continuous culture, complex medium versus minimal me-
dium, highly resolved transcriptome data versus low resolved transcriptome data
in the temporal dimension) and some mutation experiment findings.
Universality arises from the description of these very different experiments and
through the integration of transcriptome data: The evaluation of mutation ex-
periments is conveniently done by integration of new transcriptome data in the
existing model or by profile transfer as was shown for data of [
Sillers et al., 2009
].
Achieved parameters from the estimations are comparable to each other within a
20% margin and only the substrate yield required large adjustment. The model
meets one third of the proposed expectations completely, one third only partly
and it fails for the last third. To overcome the mismatch to biological findings, e.g.
the importance of butyrate up-take via the CoA-transferase pathway despite the
coupling of acetate up-take, the integration of further data is necessary, e.g. energy
metabolism data. Online annotations of energy and redox influencing metabolites
and reactions are available, for growth the stoichiometry of growth and biomass
maintenance, linear models are already developed [Papoutsakis, 1984].
4.6. FINAL CONCLUSIONS 117
Predictive Power and Usability of the Dynamical Model
The derivation of meaningful predictions is a challenging task and may often be
unsuccessful because of many parameters involved [
Jamshidi and Palsson, 2008
,
Durot et al., 2009
]. 35 Parameters, thereof 13 maximal rates are many parameters
to describe five metabolome measurement time series.
Discrepancies of the model to the experimental findings are expected for the
prediction of mutant culture behaviours, as those generally show a different
growth behaviour and different regulatory mechanisms [Tummala et al., 2003b].
It is not surprising, that all parameters, despite the reverse butyrate kinase
reaction and the ethanol dehydrogenase reaction are uncertain. This robustness is
a desired feature of a biological model [
van Riel, 2006
], however several mutation
experiments show that it is not that robust. Here the predictive power of the
model fails. This robustness can also be interpreted as high uncertainty of the
parameters, as the calculated sensitivity analyses hinted.
Still, the results of a sensitivity analysis could be transferred into an engineering
strategy if the maximal conversion velocity were directly accessible. Such an
engineering is possible but imposes several problems, as was recently shown for the
clostridial thiolase [
Mann and Luetke-Eversloh, 2013
]. Still, SA did not produce
suitable engineering strategies, since ethanol is not in the scope of this work
and also the increase of the reverse-reaction of butyrate kinase will not directly
increase butanol yields because of the present forward-reaction.
However, this model format allows a second approach that will be introduced in
the next chapter, from which other strategies can be derived.
Problems of Dynamic Models for Integration of Omics
Currently efforts are spend on creating mechanistic models of transcription and
translation via bayesian networks and differential equations [
Chen et al., 1999
,
Dhaeseleer et al., 2000
,
Arnold, 2002
,
Vijesh et al., 2013
], [
Ingalls, 2013
, chapter
7]. Implementations are reported for many networks that use Flux Balance Ana-
lysis (FBA) [
Lee et al., 2008a
,
Senger and Papoutsakis, 2008
] and many softwares
integrate FBA-methods, e.g. YanaSquare [Schwarz et al., 2007b] or CellNetAna-
lyzer [Klamt et al., 2007].
Other ways of integration transcriptome and metablome are not frequently en-
countered [
Joyce and Palsson, 2006
,
Jamshidi and Palsson, 2008
] and thereby sev-
eral problems concerning implementation and computation needed to be solved
here.
First, only two softwares were found that were able to allow integration of the
model formalism into their architecture and all following analyses needed to be
build on this integration. Missing SBML support for this model is a good indica-
tion that this model type is not widely spread, although a work-around solution
118 CHAPTER 4. AUTOMATED DYNAMIC MODEL CREATION
may be present by using SED-ML
∗
[
Waltemath et al., 2011
]. SBTOOLBOX2 was
made the tool of choice because of two reasons: Codes are publicly available and
the text-based format of the model allowed easy automation of model creation and
evaluation. This is necessary to leverage one computational problem of parameter
estimation: Stiffness of the differential equations is increased by the integration
of transcript expression profiles. In steep profiles only many small time steps of
the integrator can lead to convergence, hence a huge number of calls for each
interpolation of transcript levels is necessary. Computational speed is proportional
to these calls if the number of integrated profiles is large enough. One solution to
this problem will be presented in the next chapter. This problem and its solution
are unique to this type of model and represent a new area of research for the
modelling science.
∗Personal communication, Dagmar Waltemath, Rostock
Chapter 5
Principal Component Analysis In
Modelling
That honey is sweet I refuse to assert;
that it appears sweet I fully grant.
Timon of Phlius
Principal Component Analysis (PCA) is widely used, this chapter proposes
three independent applications to this method. After its introduction (5.1), its
historical use as compression method to alleviate simulation effort in the dynamical
model from interpolations of transcript expression data (5.2). Very closely related
to this, is a optimisation tool based on PCA to alternate transcript expression
data under the side condition of preserved dynamic (5.3). The third application
then discusses the uses of this approach for a novel clustering method (5.4). A
conclusion is given afterwards (5.5).
120 CHAPTER 5. PRINCIPAL COMPONENT ANALYSIS IN MODELLING
5.1 Introduction of PCA
Mathematical Background
An introduction to principal component analysis (PCA) and its mathematical
derivation is given by [Abdi and Williams, 2010].
The set
D
is the set of all transcript expression data in a matrix format in which
columns represent time and rows represent transcripts:
D={xj(ti), j = 1...NJ, i = 1...NT}.(5.1)
Usually
NT≪NJ
. A PCA of the data consists in a dimensionality reduction of
D
into few representative basis functions, called principal components (PC) with
respective coefficients (
c
). Each of the PCs is sorted according to their contri-
bution to the data’s variance. If indeed the data’s variance is low dimensional
then PCA is able to compress the data. It is reported that the first principal
components are usually representative for the data, they are optimal for compres-
sion [
Janes and Yaffe, 2006
]. Conversely, higher components may bear structural
information of the data [Yeung and Ruzzo, 2001].
Here, it will be assumed that only the first
NP< NT
components are relevant
and further components hold information about uncertainties. This happens on
the expense of low abundant pattern of the data that are furthermore neglected.
The basis representation of
xj
(
t
) in terms of principal components is then given
by:
xj(t) =
NP
X
p=1
cjpPCp(t) (5.2)
Uses of PCA
Besides its compression abilities [
Janes and Yaffe, 2006
], PCA has a manifold
of other uses. Perturbed states are distinguished via PCA at the metabolome
and transcriptome level from the unperturbed states [
Dutta et al., 2009
]. Re-
verse engineering approaches were possible by calculation of a PC representation
of gene expression data [
Yeung et al., 2002
]. It furthermore allows the imputa-
tion of missing data-points in a less error-prone way than regression models
[
Troyanskaya et al., 2001
]. Correlations of the data to the principal components
help in the observation of genome-wide effects [
Alter et al., 2000
]. Evaluation of
global sensitivity results was recently shown to be facilitated when the principal
components where calculated [
Sumner et al., 2012
]. The use of PCA is reported
to minimise effects of measurement noise prior to clustering in several publications
[
Brown et al., 2005
,
Janes and Yaffe, 2006
]. Other authors argue that it may de-
grade cluster quality [
Yeung and Ruzzo, 2001
]. Uses of PCA for clustering are
only reported within a side-note [Holter et al., 2000].
5.2. DATA COMPRESSION IN THE DYNAMICAL MODEL 121
5.2 Data Compression in the Dynamical Model
Preliminaries
The dynamic model of the previous chapter suffers severe difficulties when it
comes to parameter estimations. The global sensitivity analysis shows on the
one hand that the parameters are not accurately estimated. On the other hand
the criterions to stop the estimation are not met rapidly enough. Interrupting
the simulation manually results in a highly non-reproducible parameter set. In
the scope of automation and reproducibility the convergence speed is a serious
problem.
Transcript Expression Data Problems
One reason why the dynamical model is computationally demanding is the in-
tegration of highly dynamic transcript expression profiles into the equations.
Changes of expression levels in the continuous culture consist up to two orders of
magnitudes.
Noise of these data further impedes the estimation speed. It is hardly accessible
due to the missing replicates of the available data. Highly expressed genes during
the solventogenic chemostat can be averaged to estimate noise, which is around
50% on the nominal scale, or 0.5 on the log scale.
Solutions
Dynamics of data cannot be reduced, however the number of dynamic profiles can
be reduced by choosing a suitable data-model, e.g. transcript within the same
open reading frame usually have similar dynamics and levels. Such a data-model
further needs to eliminate noise from the data.
Clustering offers one possibility to reduce size of data and create regulatory as-
sumptions, as seen before (3.5).
The performance of alternative representations in terms of non-linear regression is
greatly known in the biological community. The right choice of regressors or basis
functions is a non-trivial problem that requires in-depth knowledge of underlying
data structure. Also such a representation should be useful in a greater scope
than simple representation, e.g. the unravelling of regulatory dependencies. For
automation both these facts are critical.
Other possibilities of convergence amelioration by introducing system control
theory approaches as semi-quantitative measurements, Kalman filtering or network
modular topology increase estimation quality and convergence [
van Riel, 2006
].
The principal component representation of the data is a self-contained repres-
entation of the data and does not require prior knowledge, it also discriminates
between data and data-noise.
122 CHAPTER 5. PRINCIPAL COMPONENT ANALYSIS IN MODELLING
PCA-assisted Data Compression
Equation 5.2 and the observation that usually
NP< NT≪NJ
, it is easy to
conclude that in a model in which all
NJ
transcript data levels are calculated, the
major computation time is used for the interpolation. However, all the information
necessary for calculation is stored in the
NT
components of the PC expansion.
For the CM, there are 14 different profiles and
NT
= 5, hence it is always better
to calculate only the
NT
PCs instead of all 14 profiles. The necessary 5
·
14 = 70
coefficient
cjp
can be integrated to the model prior calculation (B.3.2). The
calculation effort of the original model is
NT·NJ
= 70 versus
N2
T
= 25 of this
compressed model. This corresponds to a theoretical improvement of speed by
63%. Computing 200 independent consecutive runs of the CM and its compressed
format, took 61 sto 29 s, respectively. This is a gain of 53%.
5.3. OPTIMISATION APPROACH VIA PCA 123
5.3 Optimisation Approach via PCA
Time-resolved steering of promoter activity was a project goal within COSMIC2.
It can be expected that the directed change of transcript levels over time will be
in research focus soon. Models require to take this challenge and to guide the
experimentalist through the outcomes of different strain designs and temporal
profile designs.
The presented dynamical model contains transcript expression data and can be
used for such a prediction, if only a suitable scheme of alternation of these profiles
can be found.
Bootstrapping of Transcriptome Data Does not Influence the Metabolic Spec-
trum
The most general approach consists in the random generation of artificial tran-
script expression profiles. However, this easily results in numerical problems and
unrealistic model results.
A more directed approach than random design is toggling existing measurement
data, also known as bootstrapping [
Pattengale et al., 2010
]. The transcript ex-
pression data profile is considered a measurement curve that is the result of a
signal diverted by additive noise. For transcript expression profiles this approach
is feasible since the sample space is low populated in the temporal dimension
and an original profile can be parametrised for bootstrapping. Using again global
sensitivity analysis a huge number of model simulations and their effects were
studied. For the CM only the central point of the pH-shift data is sensitive (results
not shown). However, changes at this point showed no changes in productivity.
Directed Approach to Optimisation - Dynamic Features
Here, a novel approach is suggested. It includes an intelligent choice of profiles.
The data structure itself is not known but the use of PCA identifies the directions
of maximal variance, such direction will be called dynamic feature. Identifying the
PC with the dynamic aspects of the data [
Holter et al., 2000
] offers a possibility to
vary the dynamics of the data. This is referred to as dynamic features optimisation.
The study of combinations of dynamic features to construct new data restrains
the amount of possible profiles to only dynamically equal profiles (refer to 5.4).
Implementation
A script for alternation (B.4.1) was used to compute the new profiles. After
caluclation of the PCs, the scores are alternated by multiplication with 2, 1 or
0.5. Every combination of alternated scores gives a new profile from which a new
model is constructed and simulated.
124 CHAPTER 5. PRINCIPAL COMPONENT ANALYSIS IN MODELLING
Primary Target - Butanol Dehydrogenase
0
0
CAC3298
CAC3299
CAP0162
PC1
PC2
0
0
PC1
PC3
CAC3298
CAC3299
CAP0162
6 5.5 5 4.5 4
−2
0
2
pH
log ratio
Model Genes for Butanol Synthesis
CAC3298
CAC3299
CAP0162
sum
6 5.5 5 4.5 4
−1
−0.5
0
0.5
1
pH
log ratio
Principal Components of all Data
PC1
PC2
PC3
A
B
C1
C2
Figure 5.1: PCA Analysis of the transcript levels of the reaction r4|10.
(A): Three principal components are calculated from the entire transcriptome
data Dover the pH-range of the pH-shift.
(B): The time-courses of the three transcripts and their sum as responsible for
butanol synthesis from butyryl-CoA in the CM.
(C): PC coefficients. The position of the three transcripts in the three dimensional
PC-space shows that all transcript levels have a negative contribution of the first
PC. The second PC introduces major dynamic differences between CAC3299 and
the other transcripts. Contributions of the third PC are not as important.
From GSA it was reckoned, that butanol dehydrogenase is weakly sensitive.
This is also the primary choice of all strategies.
The PCA transformation from the entire data
D
gives the three PC, as shown
in figure 5.1,A. The profiles of the three relevant transcripts of reaction
r4|10
,
CA
C3298
, CA
C3299
, CA
P0162
add approximately to a straight line (figure 5.1,B)
- their coefficients are marked in green in figure 5.1,C1 and C2. As expected,
the profiles of CA
C3298
and CA
C3299
are very similar in their first PC. Their
major difference is found in the second PC, CA
P0162
and CA
C3298
have a negative
contribution, CA
C3299
a positive contribution. The contribution of the third PC
is nearly negligible for all three transcript levels.
Alternation of the PCA coefficients according to the implemented scheme, yields
5.3. OPTIMISATION APPROACH VIA PCA 125
a set of curves with different dynamic impacts. The maximum is attained always
at pH 4.5, the dynamics are shifted to even lower values during acidogenesis
(figure 5.2). For each profile, metabolite spectra are calculated (figures 5.3 and
5.8 5.5 5.1 4.8 4.5
−5
0
5
log ratio
5.8 5.5 5.1 4.8 4.5
−10
−5
0
pH
A
pH
B
log ratio of
transcript expression
Figure 5.2: Profile design for
r4|10
from dynamic features. Coefficients of the
individual profiles were alternated by either doubling or halving. Profiles prior to
normalisation (A) and after normalisation to the maximum (B) shows that the
maximal amount of enzyme, corresponding the maximal velocity is reached for all
profiles after the shift. Dynamics are unaltered since the profiles intersect at the
same level during the pH-shift.
5.4). Neither profile from the dynamic features increases the pull from glycolysis
to generate butanol, only butyrate is converted during acidogenesis to butanol.
Despite the earlier start of butanol synthesis in these scenarios, no further increase
of butanol yield is visible.
Other Targets
The calculation of dynamic features of butyryl-CoA synthesis from acetoacetyl-CoA
(
r34
) shows that this reaction is not limiting, alternation of the transcript profiles
does not yield an improvement. Down-regulation of the enzymes accumulates
the substrate and thereby leads to emptying the butyryl-phosphate and butyryl-
CoA pools. Interestingly, no changes on the upper-branch of acid or ethanol
synthesis is seen. Apparently, these reactions are limited. The surplus availability
of acetoacetyl-CoA leads to increased acetone formation that crashes when no
butyrate is present in the medium anymore.
The phosphtransbutyrylase reaction (
r45
) and the butyrate-kinase reaction (
r56
)
both have an identical influence. Butanol yields are increased by 15
mM
and
acetone is reduced by 10 mM at the end of the fermentation (figure 5.5).
126 CHAPTER 5. PRINCIPAL COMPONENT ANALYSIS IN MODELLING
0 50 100
0
10
20
30
Acetate
conc. [mM]
0 50 100
0
20
40
60
80 Butyrate
0 50 100
0
10
20
30 Acetone
time [h]
0 50 100
0
5
10
15 Ethanol
time [h]
conc. [mM]
0 50 100
0
10
20
30
40 Butanol
time [h]
Figure 5.3: Metabolic spectrum of the transcript optimization of r4|10.
5.3. OPTIMISATION APPROACH VIA PCA 127
0 50 100
0
0.2
0.4
AcCoA
0 50 100
0
0.02
0.04
0.06 AcAcCoA
0 50 100
0
0.5
1BuCoA
time [h]
0 50 100
0
0.2
0.4 BUP
time [h]
conc. [mM]
0 50 100
0
1
2AcAc
conc. [mM]
0 50 100
0
0.01
0.02
0.03 ACP
conc. [mM]
Figure 5.4: Metabolic spectrum of the transcript optimisation of r4|10.
128 CHAPTER 5. PRINCIPAL COMPONENT ANALYSIS IN MODELLING
40 60 80
0
30
60
time [h]
conc. [mM]
Acetate Butyrate Acetone Ethanol Butanol
40 60 80
time [h]
6 5 4
−2
0
2
pH
log. expression
original transcript
optimised transcript
AB
C
Figure 5.5: Profile design for r4|5from dynamic features.
(A): original product spectrum
(B): new product spectrum
(C): original and new time-course of CAC3076 prior to normalisation
5.4. CLUSTERING FROM PRINCIPAL COMPONENTS 129
5.4 Clustering from Principal Components
For this section it is necessary to introduce some preliminary thoughts on current
clustering techniques, because PCA will be used in a somewhat different way of
thought than current techniques.
5.4.1 Introduction
The Current Paradigm in Transcriptome Analysis
Many problems in Omics are classification problems [
Nobeli and Thornton, 2006
].
A class or a cluster is a set items that share the same properties. In the scope
of transcript level profiles this classification is based on a similarity evaluation,
like the euclidean distance of two profiles or their co-variance: From the com-
bination of the correct metric with an expectation of measurement noise, the
clustering algorithm derives groups that share a low inner-cluster variance and
a large intra-cluster variance [
Lukashin and Fuchs, 2001
]. This algorithm re-
quires user-input on how much measurement noise is expected or how many
clusters constitute the data. Through this, a cluster is a function of the chosen
metric, the chosen cluster algorithm and a set of necessary parameters. The
number of employable metrics and clustering algorithms is extremely large
[Jiang et al., 2004b,Brown et al., 2005,Janes and Yaffe, 2006].
Critique of the Current Paradigm
The current paradigm has helped in the interpretation of many data since
first large data sets became first available: The original paper by Eisen et al.
[
Eisen et al., 1998
] was cited 13587 times
∗
. However, what are the limitations of
transcriptional analysis by clustering?
It is well known, that not all transcripts are equally well transcribed during
the amplification reaction. From a stochastic point of view the multiplication
of very low abundant transcripts is a rare event because of the meeting prob-
ability of enzyme and transcript. Factors like the affinity between enzyme and
transcript, stress factors create an additional bias that is variable also in time.
This leads to the conclusion that quantities of co-transcribed genes are not equal
[You and Yin, 2000,Feder and Walser, 2005].
Further, clustering is known to reveal operons because the genes behind an operon
are concomitantly transcribed. However, metabolism-wide events will effect mul-
tiple open reading frames. These frames are not all transcribed behind promoters
of identical strength, the dynamics of expression will be similar but not the
amounts. These limitations mostly attack similarity metrics that are based on
∗15th August, 2013
130 CHAPTER 5. PRINCIPAL COMPONENT ANALYSIS IN MODELLING
distances. Correlation metrics are better suited tools regarding these limitations,
they are, however, more severely affected by measurement noise.
Finally, there is the question of information between clusters: With any metric
it is only possible to make a binary decision whether two profiles are contained
within the same cluster or not, nothing more.
5.4.2 Geometric Approach to Clustering
PCA offers a possibility to achieve a geometrical dissection of the data in a similar
way as a correlation metric but without loosing relations between the different
geometric objects, and it dissects the measurement noise.
The
NT
dimensional vector of PCA coefficients
cjp
will be furthermore named
trait of the
p
th PC to focus on the fact that the PCs are universal within the
data - they span a space - and the appearance, the phenotype of a transcript
level profile depends on the impact of each individual PC. Indeed, all possible
dynamics of the treated experiment are coded within the PCs [
Holter et al., 2000
].
Conventional Clustering of Traits
In a first attempt the use of published clustering methods like k-nearest neighbours
or hierarchical clustering to identify agglomerations of
cjp
in the spanned space
was carried out. This attempt is difficult for the beforehand mentioned reasons:
First, the determination which amount of distance is due to measurement noise can
not be carried out, because noise is encoded in the higher dimensional components
and it is not transferred to the traits. Second, the traits are sorted according
to their importance in the total data variance. This imposes the necessity to
intelligible weighing of each dimension of PCs. One such weight could be the
variance contribution of each PC to the data. Since structural information may
be present in higher components [
Yeung and Ruzzo, 2001
], and not the entire
variance is considered, this approach is discouraged here.
Clustering by Tiling of the PC-Space
In a second attempt, it seems intelligent to start with new assumptions and to
try a new approach for making sense of traits with the help of a suitable tiling
that neither requires a clustering algorithm, nor a distance metric.
The space spanned by the principal components, is the space of all achievable
dynamics CPC:
CPC := {PC1,PC2, ...PCNP}
In order to construct a tiling of this space, a meaningful concept for combination of
dynamics needs to be introduced. The trivial tiling is a partition into half-spaces
5.4. CLUSTERING FROM PRINCIPAL COMPONENTS 131
according to the sign of the respective components
†
. By pure visual inspection
any agglomeration of points within one partition may be considered a tiling, as in
figure 5.6,A. This is the established approach. Now, consider a straight line from
the origin that connects to any point. Mathematically, all points on a line in this
space do share the same proportions of each PCs.Such transcripts are called here
dynamically equal with respect to that line. In order to account for uncertainties
and to be of practical use, this line-object requires a thickness. Because there are
several possibilities to define this thickness, assumptions are required to define
this new geometric object:
1.
A maximum principle: The line-object should not be larger than its con-
taining half-space.
2.
An equality principle: Two line-objects are non overlapping and equally
sized.
3. A geometrical principle:
a) either the half-spaces are bisected into equal parts
b) or the half-spaces are trisected into equal parts.
The first assumption accounts for the impact of signs of traits on the phenotype a
change of sign may have significant effect on the overall phenotype. The second
assumption makes sure that no transcript expression profile is present in two
different tilings. The third assumption is the core of angular traits, because it
represents the human factor that classifies the relation between the coefficients
whether one is stronger present in the data than the other. Choosing a trisection
helps in classifying the traits that are strongly different to each other (sectors
1 and 3 in figure 5.6,C) by assuming that equally sized dynamic aspects are of
interest. By choosing the bisection, equal coefficients are put out of focus. Due
to the curse of dimensionality, this is the approach taken here. The number of
cones per two coefficients will be named
nangle
. For two principal components,
nangle
= 8 cones are constructed (figure 5.6, B), by increasing the number to
four principal components, this number is increasing to
nangle
= 8
3
= 512 in the
case of trisection 12 cones are constructed from two principal components, and
nangle = 123= 1728 from four principal components.
Both tilings further provides a possibility to define pairs of co-regulated and
anti-regulated transcript expression data: For a chosen cone, its point-reflection
is anti-regulated because all traits are reversed in their signs (figure 5.6,B, cones
2 and 5). A trait within such a cone will be called angular trait, it is defined as
αjp = arctancjp
cj1.
†There are 2NPhalf-spaces.
5.4. CLUSTERING FROM PRINCIPAL COMPONENTS 133
PC2
-4 -3 -2 -1 0 1 2 3 4
PC1
-5 -4 -3 -2 -1 0 1 2 3 4 5 6
PC1
PC2
PC3
PC2
-4 -3 -2 -1 0 1 2 3 4
PC1
-5 -4 -3 -2 -1 0 1 2 3 4 5 6
Figure 5.7: Pre-clustered data from cell cycle microarray data of yeast in a
PC-representation. Clusters are represented as colors. Cone-like and beam-
like structure are readily observed. Lines are manually drawn into the graphs.
Adaption from [Yeung and Ruzzo, 2001].
sectors rapidly converges close to zero (table 5.1). There is no best choice of
NP
known [Yeung and Ruzzo, 2001]. Examples of clusters are shown in figure 5.9.
Table 5.1: Overview of angular traits. The fraction of occupied to all sectors
is presented as function of
NP
and
nangle
. In the continuous culture
NT
=4,
NJ
= 3807 while in the batch culture
NT
= 25,
NJ
= 1862 and in the RT-PCR
experiment NT= 91, NJ= 181.
experiment nangle NP= 2 NP= 3 NP= 4 NP= 5 NP= 6 NP= 7
conti. cult. 8 1 0.5 0.22 - - -
batch. cult. 8 1 0.5 0.25 0.1157 0.034 0.0064
RT-PCR 8 0.5 0.2031 0.0684 0.0146 0.0026 0
Transcript data from batch culture
This data-set shows a rich dynamic. Four PCs are needed to model more than 80%
of the data and sectors are the most occupied compared to the other two data-sets.
One can expect to find many dynamic features. The approach allows to cover
dynamically very similar and very close profiles whose levels spread during the
end of fermentation when sporulation of the cells and degradation of transcripts
occur.
In figure 5.9, examples of co-regulated and anti-regulated clusters are shown in
A1 and A2 respectively.
134 CHAPTER 5. PRINCIPAL COMPONENT ANALYSIS IN MODELLING
0
20
40
60
80
100
120
continuous culture
batch culture
RT-PCR
Covered Variance
Np=6
Np=5
Np=4
Np=3
Np=2
Np=1
Figure 5.8: Variance covered by the principal component representation of tran-
scriptome experiments from continuous culture, from batch culture and from
RT-PCR data from B.subtilis.
Transcript data from continuous culture
This data-set shows a low dynamic and the smallest in temporal size. Three
components are needed to cover more than 80% of the variance. Due to the lack
of temporal resolution, clusters sizes are very large.
The example in 5.9, B shows again the co-regulated and anti-regulated clusters.
RT-PCR
This data-set is the smallest size although it has the largest resolution in time.
Clustering from PCs shows that only two of four half-spaces are occupied by the
loadings of the second PC. This takes into account that this data is only positive
and no anti-regulated clusters can be expected. Further, two components suffice
to model more than 80% of the data’s variance.
The example in 5.9,C shows the typical profiles that contain a spike.
5.5. FINAL CONCLUSIONS 135
Figure 5.9: Angular traits examples from three data-sets. Examples are drawn by
random selection from the whole data for better visualisation. Axes labels are
omitted for simplicity.
(A1 & A2): batch data, NP= 4, co-regulated and anti-regulated
(B1 & B2): continuous data, NP= 4, co-regulated and anti-regulated
(C): RT-PCR data, NP= 4
5.5 Final Conclusions
PCA seems to encounter a revival in literature
§
, methods that use PCA are widely
accepted and sophisticated alterations, e.g. rotation of components are applicable
[
Abdi and Williams, 2010
]. The compression function of PCA is greatly known
since the famous eigen-faces have been published [
Turk and Pentland, 1991
]. Re-
cently, a report was published on the use of PCA to simplify sensitivity analysis
outputs in a meaningful way [
Sumner et al., 2012
]. The here presented analyses
add three further application of PCA in the biological sciences.
§
more than 21000 articles were published as of 16th august 2013, in Google Scholar, one year
earlier only 2800 were published
136 CHAPTER 5. PRINCIPAL COMPONENT ANALYSIS IN MODELLING
First, a compression of transcript expression level profiles was calculated to in-
crease computational power of the parameter estimation. This is not treated
because such dynamic model is not published. Second, a model-assisted optimisa-
tion approach was suggested from PCA that allowed the directed alternation of
transcript expression data dynamics. This is also new because of the novel model
structure. Finally, the use of principal components for clustering was elucidated
in a hitherto untreated way for the identification of dynamically equal profiles.
A Novel Clustering Approach Was Introduced
The here presented algorithm for clustering is to best knowledge novel in literature.
The uses of PCA are manifold and its use in clustering was limited to preprocessing
of data [
Yeung and Ruzzo, 2001
]. Here, two concepts, angular traits and angular
clusters make use of the information stored in the principal components.
Four challenges were defined by [
Jiang et al., 2004b
] to a good clustering al-
gorithm:
1. No dependance on prior knowledge:
No such knowledge is necessary here and parameter numbers are low.
2. High fidelity to filter signal from noise:
Noisy components of the data are filtered by PCA in the higher dimensional
components.
3. Possibility to build hierarchical structures:
The number of sectors
nangle
and the angular closeness approach can be
used to build a hierarchical structure.
4. Possibility to retrieve relationships between clusters:
Anti-regulation and co-regulation are retrieved by comparison of sectors.
Validation of clusters is usually carried out by assessing several similarity criteria
[
Jiang et al., 2004b
], this is not possible here since by definition of sectors, genes
within one sector are considered dynamically equal. A recurrence to other similarity
metrics like the euclidean space is not possible. A second possibility is given
by comparing the given clustering with a master clustering and to compare the
sorting of both. One similar method to this clustering is the Pearson correlation
to assess co-expression [
Carrera et al., 2009
], the building of a coherent cluster
[
Jiang et al., 2004a
] using Pearson correlation requires the definition of a threshold.
Such definition was avoided here by introducing a qualitative factor, the geometric
assumptions.
5.5. FINAL CONCLUSIONS 137
Not All Data is Equally Well Represented by PCA
The use of PCA in knowledge generation is known [
Alter et al., 2000
], the here
presented algorithm can be considered as a complementary approach to already
existing ones, because it emphasises on structure evaluation. Since the number of
considered principal components is directly related to the structure properties,
angular traits becomes difficult when a large number of components is possible
but only few data are present (
NT≈NJ
), which is the case for the RT-PCR data.
Data curves that are degenerated in the sense of having a unique profile compared
to the others are passing through undistinguished when
NP
is chosen too small.
When in contrast in such data
NP
is chosen too large, major pattern are split
into small pieces. The here presented algorithm works under the assumption that
all noise is equally spread in components with numbers larger than
NP
. In the
RT-PCR data this is not true. An identification of the correct
NP
would require
to measure the information content of angular traits per dimensions. Ultimately,
this leads to the question, if the proposed equality principle must be relaxed.
PCA for Modelling
The use of PCA for enhancing calculation speed by reducing interpolation effort
was proven. Its use for optimisation also proposes beneficial outcomes. Here, the
effect of such a change is critical. Large changes of the solvent spectrum lead to
large effect on transcriptome [
Tummala et al., 2003a
], the change of transcript
expression pattern should be soft. In order to investigate the effect of these
changes further, more experimental data is required.
Summary
1.
Workflows were programmed that allowed automatic harvesting of the
KEGG-database. Compound information, Pfam-motif annotation of various
organisms, gene-enzyme-reaction-reaction pair mapping were achieved and
created the basis for automated creation of static and dynamic models.
2.
A formalism for the integration of pathway information and transcript ex-
pression data was proposed. Transcriptome data was organised in a coherent
way and it was visualised successfully. Inspection of the pathway models and
of comparative maps to B. subtilis allowed hypothesis generation of a novel
pathway that requires the activity of an unannotated 3-hydroxybutyrate
dehydrogenase.
3.
Application of the domain-grammar hypothesis inspired the formulation
of an algorithm for Pfam-motif collection of 3-HBDH from 750 organisms
from a frequentist point of view. Integration of experimental data by
assuming parallel regulation in clusters constructed hypotheses that can be
experimentally verified. The ranking of hypotheses was enabled by weighting
the results. This weight needs re-thinking in order to facilitate the ranking
procedure. The CA
C3335
gene was suggested to contain a 3-HBDH activity
during solventogenesis.
4.
Automated creation and integration of transcript data into KEGG-derived
dynamic models was programmed. This model type succeeded the repres-
entation of the pH-shift experiment without containing a pH-dependency.
Cross-validation between batch and continuous data was successful and the
parametrised model was qualitatively validated by simulating published
mutation experiments. Parameter certainty was discussed and possible
extensions of the model were proposed.
5.
Local and global sensitivity analysis were performed for bottleneck identi-
fication. GSA showed that butanol dehydrogenase cannot pull enough the
carbon from other reactions like the butyrate production that was shown to
139
140 CHAPTER 5. PRINCIPAL COMPONENT ANALYSIS IN MODELLING
be cycling between the kinase and the CoA-transferase pathway. The model
was shown to be robust against the major parameter changes.
6.
Speed optimisation of the dynamic model was performed using PCA. This ap-
proach inspired a novel algorithm for model evaluation based on the dynamics
of transcript level data. An optimal transcript profile for butanol dehydro-
genase could not be found from the data but for butyrate transacetylase.
7.
Furthermore, a clustering algorithm was constructed from PCA by chan-
ging the perspective from parametric clustering to geometric assumptions.
The characterisation of data according to their angular traits and angular
similarity gives a promising novel way of performing informative clustering.
Co-regulated and anti-regulated genes are easily computed by this algorithm.
Bibliography
[Abdi and Williams, 2010]
Abdi, H. and Williams, L. (2010). Principal component
analysis. Wiley Interdisciplinary Reviews: Computational Statistics, 2(4):433–59.
[cited at p. 120, 135]
[Agren et al., 2013]
Agren, R., Liu, L., Shoaie, S., Vongsangnak, W., Nookaew, I., and
Nielsen, J. (2013). The raven toolbox and its use for generating a genome-scale
metabolic model for penicillium chrysogenum. PLoS Comput Biol, 9(3):e1002980.
[cited at p. 29, 116]
[Aittokallio and Schwikowski, 2006]
Aittokallio, T. and Schwikowski, B. (2006). Graph-
based methods for analysing networks in cell biology. Briefings in Bioinformatics,
7(3):243–55. [cited at p. 32, 34, 78]
[Akesson et al., 2004]
Akesson, M., F¨orster, J., and Nielsen, J. (2004). Integration of gene
expression data into genome-scale metabolic models. Metabolic engineering, 6(4):285.
[cited at p. 36]
[Aksenov et al., 2005]
Aksenov, S., Church, B., Dhiman, A., Georgieva, A., Sarangapani,
R., Helmlinger, G., and Khalil, I. (2005). An integrated approach for inference and
mechanistic modeling for advancing drug development. FEBS Letters, 579(8):1878 –
83. [cited at p. 3]
[Alsaker and Papoutsakis, 2005]
Alsaker, K. and Papoutsakis, E. (October 15, 2005).
Transcriptional program of early sporulation and stationary-phase events in clostridium
acetobutylicum. Journal of Bacteriology, 187(20):7103–18. [cited at p. 21]
[Alsaker et al., 2010]
Alsaker, K., Paredes, C., and Papoutsakis, E. (2010). Metabolite
stress and tolerance in the production of biofuels and chemicals: Gene-expression-based
systems analysis of butanol, butyrate, and acetate stresses in the anaerobe clostridium
acetobutylicum. Biotechnology and Bioengineering, 105(6):1131–47.
[cited at p. 21, 22,
35, 39, 46]
[Alsaker et al., 2004]
Alsaker, K., Spitzer, T., and Papoutsakis, E. (2004). Transcriptional
analysis of spo0a overexpression in clostridium acetobutylicum and its effect on the
cell’s response to butanol stress. Journal of Bacteriology, 186(7):1959–71.
[cited at p. 14,
20, 21]
141
142 BIBLIOGRAPHY
[Alter et al., 2000]
Alter, O., Brown, P., and Botstein, D. (2000). Singular value decom-
position for genome-wide expression data processing and modeling. Proceedings of the
National Academy of Sciences, 97(18):10101–6. [cited at p. 120, 137]
[Altschul et al., 1997]
Altschul, S., Madden, T., Schaeffer, A., Zhang, J., Zhang, Z., Miller,
W., and Lipman, D. (1997). Gapped blast and psi-blast: a new generation of protein
database search programs. Nucleic Acids Research, 25(17):3389–402. [cited at p. 64]
[Alves et al., 2006]
Alves, R., Antunes, F., and Salvador, A. (2006). Tools for kinetic
modeling of biochemical networks. Nature biotechnology, 24(6):667–72.
[cited at p. 3, 92]
[Amador-Noguez et al., 2011]
Amador-Noguez, D., Brasg, I. A., Feng, X., Roquet,
N., and Rabinowitz, J. D. (2011). Metabolome remodeling during the acidogenic-
solventogenic transition in clostridium acetobutylicum. Applied and Environmental
Microbiology, 77(22):7984–97. [cited at p. 17, 20, 95]
[Andrade and Vasconcelos, 2003]
Andrade, J. C. and Vasconcelos, I. (2003). Continu-
ous cultures of clostridium acetobutylicum: culture stability and low-grade glycerol
utilisation. Biotechnology Letters, 25:121–5. [cited at p. 16]
[Arnold, 2002]
Arnold, S. (2002). Kinetic Modelling of Gene Expression. PhD thesis,
University of Stuttgart, Germany. [cited at p. 89, 117]
[Badr et al., 2001]
Badr, H., Toledo, R., and Hamdy, M. (2001). Continuous acetone-
ethanol-butanol fermentation by immobilized cells of clostridium acetobutylicum. Bio-
mass and Bioenergy, 20(2):119 – 32. [cited at p. 18]
[Bahl et al., 1982a]
Bahl, H., Andersch, W., Braun, K., and Gottschalk, G. (1982a).
Effect of ph and butyrate concentration on the production of acetone and butanol by
clostridium acetobutylicum grown in continuous culture. Applied Microbiology and
Biotechnology, 14:17–20. [cited at p. 13, 16]
[Bahl et al., 1982b]
Bahl, H., Andersch, W., and Gottschalk, G. (1982b). Continuous
production of acetone and butanol by clostridium acetobutylicum in a two-stage
phosphate limited chemostat. Applied Microbiology and Biotechnology, 15:201–5.
[cited at p. 16]
[Bajad et al., 2006]
Bajad, S., Lu, W., Kimball, E., Yuan, J., Peterson, C., and Ra-
binowitz, J. (2006). Separation and quantitation of water soluble cellular metabolites
by hydrophilic interaction chromatography-tandem mass spectrometry. Journal of
Chromatography A, 1125(1):76 – 88. [cited at p. 20]
[Balodimos et al., 1988]
Balodimos, I., Kashket, E., and Rapaport, E. (1988). Metabolism
of adenylylated nucleotides in clostridium acetobutylicum. Journal of Bacteriology,
170(5):2301–2305. [cited at p. 17]
[Barab´asi and Oltvai, 2004]
Barab´asi, A.-L. and Oltvai, Z. N. (2004). Network biology:
understanding the cell’s functional organization. Nature Review Bioinformatics, 5:101–
13. [cited at p. 32]
[Barbeau et al., 1988]
Barbeau, J., Marchal, R., and Vandecasteele, J. (1988). Con-
ditions promoting stability of solventogenesis or culture degeneration in continuous
BIBLIOGRAPHY 143
fermentations of clostridium acetobutylicum. Applied Microbiology and Biotechnology,
29(5):447–55. [cited at p. 10]
[Barker et al., 1978]
Barker, H., Jeng, I., Neff, N., Robertson, J., Tam, F., and Hosaka, S.
(1978). Butyryl-coa:acetoacetate coa-transferase from a lysine-fermenting clostridium.
Journal of Biological Chemistry, 253(4):1219–25. [cited at p. 66]
[Bassett Jr et al., 1999]
Bassett Jr, D., Eisen, M., and Boguski, M. (1999). Gene expres-
sion informatics its all in your mine. Nature Reviews, 21:51–5. [cited at p. 37]
[Beesch, 1952]
Beesch, S. (1952). Acetone-butanol fermentation of sugars. Industrial &
Engineering Chemistry, 44(7):1677–82. [cited at p. 8, 16]
[Bentele et al., 2004]
Bentele, M., Lavrik, I., Ulrich, M., Ster, S., Heermann, D., Kalthoff,
H., Krammer, P., and Eils, R. (2004). Mathematical modeling reveals threshold
mechanism in cd95-induced apoptosis. The Journal of Cell Biology, 166(6):839–51.
[cited at p. 106, 110]
[Bonnet et al., 2013]
Bonnet, E., Calzone, L., Rovera, D., Stoll, G., Barillot, E., and
Zinovyev, A. (2013). Binom 2.0, a cytoscape plugin for accessing and analyzing
pathways using standard systems biology formats. BMC Systems Biology, 7(1):18.
[cited at p. 3]
[Borden and Papoutsakis, 2007]
Borden, J. and Papoutsakis, E. (2007). Dynamics of
genomic-library enrichment and identification of solvent tolerance genes for clostridium
acetobutylicum. Applied and Environmental Microbiology, 73(9):3061–8.
[cited at p. 18]
[Bowles and Ellefson, 1985]
Bowles, L. and Ellefson, W. (1985). Effects of butanol on
clostridium acetobutylicum. Applied and Environmental Microbiology, 50(5):1165–70.
[cited at p. 17]
[Boynton et al., 1994]
Boynton, Z. L., B, G. N., and Rudolph, F. B. (1994). Intracellular
concentrations of coenzyme a and its derivatives from clostridium acetobutylicum atcc
824 and their roles in enzyme regulation. Applied and Environmental Microbiology,
60(1):39–44. [cited at p. 13, 95]
[Brekke, 2007]
Brekke, K. (2007). Butanol - an energy alternative. Ethanol Today, pages
36–9. [cited at p. 7]
[Brohee and van Helden, 2006]
Brohee, S. and van Helden, J. (2006). Evaluation of
clustering algorithms for protein-protein interaction networks. BMC bioinformatics,
7(1):488. [cited at p. 76]
[Brown et al., 2005]
Brown, M., Dunn, W., Ellis, D., Goodacre, R., Handl, J., Knowles,
J., O’Hagan, S., Spasic, I., and Kell, D. (2005). A metabolome pipeline: from concept
to data to knowledge. Metabolomics, 1(1):39–51. [cited at p. 120, 129]
[Brown et al., 2000]
Brown, M., Grundy, W., Lin, D., Cristianini, N., Sugnet, C., Furey,
T., Ares, M., and Haussler, D. (2000). Knowledge-based analysis of microarray gene
expression data by using support vector machines. Proceedings of the National Academy
of Sciences, 97(1):262–7. [cited at p. 76]
144 BIBLIOGRAPHY
[Buday et al., 1990]
Buday, Z., Linden, J., and Karim, M. (1990). Improved acetone-
butanol fermentation analysis using subambient hplc column temperature. Enzyme
and Microbial Technology, 12(1):24 – 7. [cited at p. 20]
[Cacuci, 2003]
Cacuci, D. G. (2003). Sensitivity and uncertainty analysis. Chapman &
Hall. [cited at p. 106]
[Cakir et al., 2006]
Cakir, T., Raosaheb Patil, K., Ilsen
¨
Onsan, Z.,
¨
Ozergin
¨
Ulgen, K.,
Kirdar, B., and Nielsen, J. (2006). Integration of metabolome data with metabolic
networks reveals reporter reactions. Molecular Systems Biology, 2:1–11.
[cited at p. 37,
53]
[Cammer et al., 2003]
Cammer, S., Hoffman, B., Speir, J., Canady, M., Nelson, M.,
Knutson, S., Gallina, M., Baxter, S., and Fetrow, J. (2003). Structure-based active
site profiles for genome analysis and functional family subclassification. Journal of
Molecular Biology, 334(3):387 – 401. [cited at p. 75]
[Carrera et al., 2009]
Carrera, J., Rodrigo, G., and Jaramillo, A. (2009). Model-
based redesign of global transcription regulation. Nucleic acids research, 37(5):1–11.
[cited at p. 81, 136]
[Chang et al., 2009]
Chang, A., Scheer, M., Grote, A., Schomburg, I., and Schomburg,
D. (2009). Brenda, amenda and frenda the enzyme information system: new content
and tools in 2009. Nucleic Acids Research, 37(suppl 1):D588–D592. [cited at p. 38]
[Chen et al., 1999]
Chen, T., He, H., and Church, G. (1999). Modeling gene expression
with differential equations. In Pacific symposium on biocomputing, number 29 in 4,
page 4. [cited at p. 89, 117]
[Chevenet et al., 2006]
Chevenet, F., Brun, C., Banuls, A., Jacq, B., and Christen, R.
(2006). Treedyn: towards dynamic graphics and annotations for analyses of trees. BMC
Bioinformatics, 7(1):439. [cited at p. 55]
[Cho et al., 2012]
Cho, D., Shin, S., and Kim, Y. (2012). Effects of acetic and formic
acid on abe production by clostridium acetobutylicum and clostridium beijerinckii.
Biotechnology and Bioprocess Engineering, 17:270–5. [cited at p. 13]
[Cho et al., 2003]
Cho, K., Shin, S., Kolch, W., and Wolkenhauer, O. (2003). Experi-
mental design in systems biology, based on parameter sensitivity analysis using a monte
carlo method: A case study for the tnf
α
-mediated nf-
κ
b signal transduction pathway.
Simulation, 79(12):726–9. [cited at p. 106]
[Clarke et al., 1988]
Clarke, K., Hansford, G., and Jones, D. (1988). Nature and signific-
ance of oscillatory behavior during solvent production by clostridium acetobutylicum in
continuous culture. Biotechnology and Bioengineering, 32(4):538–44. [cited at p. 14, 95]
[Company, 2006]
Company, T. D. C. (2006). Product safety assessment n-butanol. Inter-
net. Form No. 233-00247-KC-0406. [cited at p. 6]
[Covert et al., 2001]
Covert, M., Schilling, C., and Palsson, B. (2001). Regulation of
gene expression in flux balance models of metabolism. Journal of Theoretical Biology,
213(1):73 – 88. [cited at p. 36, 81]
BIBLIOGRAPHY 145
[Crown et al., 2011]
Crown, S., Indurthi, D., Ahn, S., Choi, J., Papoutsakis, E., and
Antoniewicz, M. (2011). Resolving the tca cycle and pentose-phosphate pathway
of clostridium acetobutylicum atcc 824: Isotopomer analysis, in vitro activities and
expression analysis. Biotechnology journal, 6(3):300–5. [cited at p. 56]
[Dallas et al., 2005]
Dallas, P., Gottardo, N., Firth, M., Beesley, A., Hoffmann, K., Terry,
P., Freitas, J., Boag, J., Cummings, A., and Kees, U. (2005). Gene expression levels
assessed by oligonucleotide microarray analysis and quantitative real-time rt-pcr - how
well do they correlate? BMC Genomics, 6(1):59. [cited at p. 21]
[Daub et al., 2003]
Daub, C., Kloska, S., and Selbig, J. (2003). Metagenealyse: analysis
of integrated transcriptional and metabolite data. Bioinformatics, 19(17):2332–3.
[cited at p. 90, 173]
[Dennis et al., 2003]
Dennis, G., Sherman, B., Hosack, D., Yang, J., Gao, W., Lane,
H. C., and Lempicki, R. (2003). David: Database for annotation, visualization, and
integrated discovery. Genome Biology, 4(5):P3. [cited at p. 55]
[Desai et al., 1999]
Desai, R. P., Harris, L., Welker, N. E., and Papoutsakis, E. (1999).
Metabolic flux analysis elucidates the importance of the acid-formation pathways in
regulating solvent production by clostridium acetobutylicum. Metabolic Engineering,
1(3):206 – 13. [cited at p. 13, 80, 95]
[Dhaeseleer et al., 2000]
Dhaeseleer, P., Liang, S., and Somogyi, R. (2000). Genetic
network inference: from co-expression clustering to reverse engineering. Bioinformatics,
16(8):707–26. [cited at p. 37, 64, 76, 117]
[Doris de Guzman, 2011]
Doris de Guzman, I. (2011). Green chemicals: Growing number
of chemical firms enter bio-butanol space. Internet. [cited at p. 7]
[Downer et al., 2006]
Downer, J., Sevinsky, J., Ahn, N., Resing, K., and Betterton, M.
(2006). Incorporating expression data in metabolic modeling: A case study of lactate
dehydrogenase. Journal of Theoretical Biology, 240(3):464 – 74. [cited at p. 81]
[Duarte et al., 2007]
Duarte, N., Becker, S., Jamshidi, N., Thiele, I., Mo, M., Vo, T.,
Srivas, R., and Palsson, B. (2007). Global reconstruction of the human metabolic
network based on genomic and bibliomic data. Proceedings of the National Academy
of Sciences, 104(6):1777–82. [cited at p. 36]
[Dudley, 2012a]
Dudley, B. (2012a). BP energy outlook 2030. Technical report, British
Petroleum. [cited at p. 7]
[Dudley, 2012b]
Dudley, B. (2012b). BP statistical review of world energy june 2012.
Technical report, British Petroleum. [cited at p. 6]
[Duerre, 2005]
Duerre, P. (2005). Handbook On Clostridia. Taylor & Francis.
[cited at p. 10,
11, 13, 15, 66]
[Duerre, 2007]
Duerre, P. (2007). Biobutanol: An attractive biofuel. Biotechnol. J,
2:1525–34. [cited at p. 6, 7]
[Duerre et al., 2002]
Duerre, P., Boehringer, M., Nakotte, S., Schaffer, S., Thormann, K.,
and Zickner, B. (2002). Transcriptional regulation of solventogenesis in clostridium acet-
obutylicum. Journal of molecular microbiology and biotechnology, 4(3):295.
[cited at p. 14]
146 BIBLIOGRAPHY
[Durot et al., 2009]
Durot, M., Bourguignon, P., and Schachter, V. (2009). Genome-scale
models of bacterial metabolism: reconstruction and applications. FEMS microbiology
reviews, 33(1):164–90. [cited at p. 28, 56, 78, 82, 117]
[Dusseaux et al., 2013]
Dusseaux, S., Croux, C., Soucaille, P., and Meynial-Salles, I.
(2013). Metabolic engineering of clostridium acetobutylicum atcc 824 for the high-yield
production of a biofuel composed of an isopropanol/butanol/ethanol mixture. Metabolic
Engineering, 18(0):1 – 8. [cited at p. 15]
[Dutta et al., 2009]
Dutta, B., Kanani, H., Quackenbush, J., and Klapa, M. (2009). Time-
series integrated omic analyses to elucidate short-term stress-induced responses in plant
liquid cultures. Biotechnology and bioengineering, 102(1):264–79. [cited at p. 34, 120]
[Eckert and Schuegerl, 1987]
Eckert, G. and Schuegerl, K. (1987). Continuous acetone-
butanol production with direct product removal. Applied Microbiology and Biotechno-
logy, 27:221–8. [cited at p. 18]
[Eisen et al., 1998]
Eisen, M., Spellman, P., Brown, P., and Botstein, D. (1998). Cluster
analysis and display of genome-wide expression patterns. Proceedings of the National
Academy of Sciences, 95(25):14863–8. [cited at p. 129]
[Ezeji and Blaschek, 2008]
Ezeji, T. and Blaschek, H. P. (2008). Fermentation of dried
distillers grains and solubles (ddgs) hydrolysates to solvents and value-added products
by solventogenic clostridia. Bioresource Technology, 99(12):5232 – 42. [cited at p. 17]
[Ezeji et al., 2010]
Ezeji, T., Milne, C., Price, N., and Blaschek, H. (2010). Achievements
and perspectives to overcome the poor solvent resistance in acetone and butanol-
producing microorganisms. Applied Microbiology and Biotechnology, 85(6):1697–712.
[cited at p. 18]
[Farrugia and Quigley, 2011]
Farrugia, M. and Quigley, A. (2011). Effective temporal
graph layout: A comparative study of animation versus static display methods. In-
formation Visualization, 10(1):47–64. [cited at p. 55]
[Feder and Walser, 2005]
Feder, M. and Walser, J. (2005). The biological limitations of
transcriptomics in elucidating stress and stress responses. Journal of Evolutionary
Biology, 18(4):901–10. [cited at p. 129]
[Fiehn, 2001]
Fiehn, O. (2001). Combining genomics, metabolome analysis, and bio-
chemical modelling to understand metabolic networks. Comparative and Functional
Genomics, 2(3):155–68. [cited at p. 20]
[Flannery, 2007]
Flannery, W. P. S. T. W. V. B. (2007). Numerical Recipies in C.
Cambridge University Press; 3 edition. [cited at p. 94]
[Fond et al., 1986]
Fond, O., Engasser, J., Matta-El-Amouri, G., and Petitdemange,
H. (1986). The acetone butanol fermentation on glucose and xylose. ii regulation
and kinetics in fed-batch cultures. Biotechnology and Bioengineering, 28(2):167–75.
[cited at p. 16]
[Fond et al., 1984]
Fond, O., Petitdemange, E., Petitdemange, H., and Gay, R. (1984).
Effect of glucose flow on the acetone butanol fermentation in fed batch culture. Bio-
technology Letters, 6(1):13–8. [cited at p. 20]
BIBLIOGRAPHY 147
[Forsberg, 1987]
Forsberg, C. W. (1987). Production of 1,3-propanediol from glycerol by
clostridium acetobutylicum and other clostridium species. Applied and Environmental
Microbiology, 53(4):639–43. [cited at p. 16]
[Forslund, 2011]
Forslund, K. (2011). The relationship between orthology, protein do-
main architectue and protein function. PhD thesis, Stockholm University, Sweden.
[cited at p. 56, 64, 74]
[Freeman et al., 2007]
Freeman, R., Goldovsky, L., Brosch, M., van Dongen, S., Maziere,
P., Grocock, R. J., Freilich, S., Thornton, J., and Enright, A. (2007). Construction,
visualisation, and clustering of transcription networks from microarray expression data.
PLoS Comput Biol, 3(10):e206. [cited at p. 36, 37, 76]
[Gama-Castro et al., 2008]
Gama-Castro, S., Jimnez-Jacinto, V., Peralta-Gil, M., Santos-
Zavaleta, A., Penaloza-Spinola, M. I., Contreras-Moreira, B., Segura-Salazar, J., Muniz-
Rascado, L., Martinez-Flores, I., Salgado, H., Bonavides-Martinez, C., Abreu-Goodger,
C., Rodriguez-Penagos, C., Miranda-Rios, J., Morett, E., Merino, E., Huerta, A.,
Trevino-Quintanilla, L., and Collado-Vides, J. (2008). Regulondb (version 6.0): gene
regulation model of escherichia coli k-12 beyond transcription, active (experimental)
annotated promoters and textpresso navigation. Nucleic Acids Research, 36(suppl
1):D120–D124. [cited at p. 54]
[Gheshlaghi, 2009]
Gheshlaghi, R. (2009). Metabolic pathways of clostridia for producing
butanol. Biotechnology advances, 27(6):764–81. [cited at p. 12, 17, 82]
[Golding et al., 2005]
Golding, I., Paulsson, J., Zawilski, S., and Cox, E. (2005). Real-
time kinetics of gene activity in individual bacteria. Cell, 123(6):1025–36.
[cited at p. 37,
53, 89]
[Gottwald and Gottschalk, 1985]
Gottwald, M. and Gottschalk, G. (1985). The internal
ph of clostridium acetobutylicum and its effect on the shift from acid to solvent
formation. Archives of Microbiology, 143(1):42–6. [cited at p. 13]
[G¨otz and Reuss, 2009]
G¨otz, P. and Reuss, M. (2009). Generation of regulatory hypo-
theses by descriptive modeling and data integration. New Biotechnology, 25, Supple-
ment(0):S338. ¡ce:title¿Abstracts of the 14th European Congress on BiotechnologyBar-
celona, Spain 1316 September, 2009¡/ce:title¿. [cited at p. 81]
[Green et al., 1996]
Green, E., Boynton, Z., Harris, L., Rudolph, F., Papoutsakis, E.,
and Bennett, G. (1996). Genetic manipulation of acid formation pathways by gene
inactivation in clostridium acetobutylicum atcc 824. Microbiology, 142(8):2079–86.
[cited at p. 14, 20, 65, 66, 95, 103]
[Grimmler et al., 2011]
Grimmler, C., Janssen, H., Krausse, D., Fischer, R., Bahl, H.,
Duerre, P., Liebl, W., and Ehrenreich, A. (2011). Genome-wide gene expression
analysis of the switch between acidogenesis and solventogenesis in continuous cultures
of clostridium acetobutylicum. J Mol Microbiol Biotechnol, 20:1–15.
[cited at p. 10, 22,
39, 47, 66, 68, 93, 95]
[Grupe and Gottschalk, 1992]
Grupe, H. and Gottschalk, G. (1992). Physiological events
in clostridium acetobutylicum during the shift from acidogenesis to solventogenesis
148 BIBLIOGRAPHY
in continuous culture and presentation of a model for shift induction. Applied and
Environmental Microbiology, 58(12):3896–902. [cited at p. 13, 95, 104]
[Gustafsson et al., 2009]
Gustafsson, M., H¨ornquist, M., Lundstr¨om, J., Bj¨orkegren,
J., and Tegner, J. (2009). Reverse engineering of gene networks with lasso and
nonlinear basis functions. Annals of the New York Academy of Sciences, 1158(1):265–
75. [cited at p. 88]
[Harris et al., 2000]
Harris, L., Desai, R., Welker, N., and Papoutsakis, E. (2000). Char-
acterization of recombinant strains of the clostridium acetobutylicum butyrate kinase
inactivation mutant: Need for new phenomenological models for solventogenesis and
butanol inhibition? Biotechnology and Bioengineering, 67(1):1–11. [cited at p. 14, 20]
[Harris et al., 2002]
Harris, L., Welker, N. E., and Papoutsakis, E. (2002). Northern,
morphological, and fermentation analysis of spo0a inactivation and overexpression
in clostridium acetobutylicum atcc 824. Journal of Bacteriology, 184(13):3586–97.
[cited at p. 14]
[Hartmanis and Gatenbeck, 1984]
Hartmanis, M. and Gatenbeck, S. (1984). Intermediary
metabolism in clostridium acetobutylicum: Levels of enzymes involved in the formation
of acetate and butyrate. Applied and Environmental Microbiology, 47(6):1277–83.
[cited at p. 95]
[Hartmanis et al., 1984]
Hartmanis, M., Klason, T., and Gatenbeck, S. (1984). Up-
take and activation of acetate and butyrate in clostridium acetobutylicum. Applied
Microbiology and Biotechnology, 20(1):66–71. [cited at p. 66]
[Haus et al., 2011]
Haus, S., Jabbari, S., Millat, T., Janssen, H., Fischer, R., Bahl, H.,
King, J., and Wolkenhauer, O. (2011). A systems biology approach to investigate the
effect of ph-induced gene regulation on solvent production by clostridium acetobutylicum
in continuous culture. BMC Systems Biology, 5(1):10. [cited at p. 13, 81]
[Heap et al., 2007]
Heap, J., Pennington, O., Cartman, S., Carter, G., and Minton, N.
(2007). The clostron: A universal gene knock-out system for the genus clostridium.
Journal of Microbiological Methods, 70(3):452 – 464. [cited at p. 14]
[Hernandez, 2004]
Hernandez, O. (2004). n-butyl alcohol. Technical report, UNEP.
[cited at p. 7]
[Holter et al., 2000]
Holter, N., Mitra, M., Maritan, A., Cieplak, M., Banavar, J., and
Fedoroff, N. (2000). Fundamental patterns underlying gene expression profiles: simpli-
city from complexity. Proceedings of the National Academy of Sciences, 97(15):8409–14.
[cited at p. 120, 123, 130, 132]
[Huang et al., 2009]
Huang, D., Sherman, B., and Lempicki, R. (2009). Bioinformatics
enrichment tools: paths toward the comprehensive functional analysis of large gene
lists. Nucleic Acids Research, 37(1):1–13. [cited at p. 53, 55]
[Huang et al., 1985]
Huang, L., Gibbins, L., and Forsberg, C. (1985). Transmembrane ph
gradient and membrane potential in clostridium acetobutylicum during growth under
acetogenic and solventogenic conditions. Applied and Environmental Microbiology,
50(4):1043–7. [cited at p. 13, 17]
BIBLIOGRAPHY 149
[Hucka et al., 2003]
Hucka, M., Finney, A., Sauro, H., Bolouri, H., Doyle, J., Kitano, H.,
the rest of the SBML Forum:, Arkin, A., Bornstein, B., Bray, D., Cornish-Bowden,
A., Cuellar, A., Dronov, S., Gilles, E., Ginkel, M., Gor, V., Goryanin, I., Hedley, W.,
Hodgman, T., Hofmeyr, J., Hunter, P., Juty, N., Kasberger, J., Kremling, A., Kummer,
U., Le Novere, N., Loew, L., Lucio, D., Mendes, P., Minch, E., Mjolsness, E., Nakayama,
Y., Nelson, M., Nielsen, P., Sakurada, T., Schaff, J., Shapiro, B., Shimizu, T., Spence,
H., Stelling, J., Takahashi, K., Tomita, M., Wagner, J., and Wang, J. (2003). The
systems biology markup language (sbml): a medium for representation and exchange
of biochemical network models. Bioinformatics, 19(4):524–531. [cited at p. 3]
[Huesemann and Papoutsakis, 1988]
Huesemann, M. and Papoutsakis, E. (1988). Solvent-
ogenesis in clostridium acetobutylicum fermentations related to carboxylic acid and
proton concentrations. Biotechnology and Bioengineering, 32(7):843–52. [cited at p. 13]
[Huesemann and Papoutsakis, 1990]
Huesemann, M. and Papoutsakis, E. (1990). Ef-
fects of propionate and acetate additions on solvent production in batch cultures of
clostridium acetobutylicum. Applied and Environmental Microbiology, 56(5):1497–500.
[cited at p. 12]
[Hull et al., 2006]
Hull, D., Wolstencroft, K., Stevens, R., Goble, C., Pocock, M., Li, P.,
and Oinn, T. (2006). Taverna: a tool for building and running workflows of services.
Nucleic Acids Research, 34(suppl 2):W729–W732. [cited at p. 29, 167]
[Ingalls, 2013]
Ingalls, B. (2013). Mathematical Modelling in Systems Biology: An Intro-
duction. Internet. [cited at p. 117]
[Inui et al., 2008]
Inui, M., Suda, M., Kimura, S., Yasuda, K., Suzuki, H., Toda, H.,
Yamamoto, S., Okino, S., Suzuki, N., and Yukawa, H. (2008). Expression of clostridium
acetobutylicum butanol synthetic genes in escherichia coli. Applied Microbiology and
Biotechnology, 77(6):1305–16. [cited at p. 15]
[Izak et al., 2008]
Izak, P., Schwarz, K., Ruth, W., Bahl, H., and Kragl, U. (2008).
Increased productivity of clostridium acetobutylicum fermentation of acetone, butanol,
and ethanol by pervaporation through supported ionic liquid membrane. Applied
Microbiology and Biotechnology, 78:597–602. [cited at p. 18]
[Jamshidi and Palsson, 2008]
Jamshidi, N. and Palsson, B. (2008). Formulating genome-
scale kinetic models in the post-genome era. Molecular systems biology, 4(1).
[cited at p. 117]
[Janes and Yaffe, 2006]
Janes, K. and Yaffe, M. (2006). Data-driven modelling of
signal-transduction networks. Nature Reviews Molecular Cell Biology, 7(11):820–28.
[cited at p. 120, 129]
[Jang et al., 2012]
Jang, Y., Lee, J., Lee, J., Park, J., Im, J., Eom, M., Lee, J., Lee,
S., Song, H., Cho, J., Seung, D., and Lee, S. (2012). Enhanced butanol production
obtained by reinforcing the direct butanol-forming route in clostridium acetobutylicum.
mBio, 3(5). [cited at p. 15]
[Janssen et al., 2010]
Janssen, H., Doering, C., Ehrenreich, A., Voigt, B., Hecker, M.,
Bahl, H., and Fischer, R. (2010). A proteomic and transcriptional view of acidogenic
150 BIBLIOGRAPHY
and solventogenic steady-state cells of clostridium acetobutylicum in a chemostat
culture. Applied Microbiology and Biotechnology, 87(6):2209–26. [cited at p. 13, 21]
[Janssen et al., 2012]
Janssen, H., Grimmler, C., Ehrenreich, A., Bahl, H., and Fischer, R.
(2012). A transcriptional study of acidogenic chemostat cells of clostridium acetobutyl-
icum - solvent stress caused by a transient n-butanol pulse. Journal of Biotechnology,
161(3):354 – 65. [cited at p. 17, 21]
[Jarzebski et al., 1992]
Jarzebski, A., Goma, G., and Soucaille, P. (1992). Modelling
of continuous acetonobutylic fermentation. Bioprocess and Biosystems Engineering,
7:357–361. [cited at p. 80]
[Jiang et al., 2004a]
Jiang, D., Pei, J., Ramanathan, M., Tang, C., and Zhang, A. (2004a).
Mining coherent gene clusters from gene-sample-time microarray data. In Proc.
ACM SIGKDD Int. Conf. Knowledge Discovery and Data Mining. University, West.
[cited at p. 136]
[Jiang et al., 2004b]
Jiang, D., Tang, C., and Zhang, A. (2004b). Cluster analysis for
gene expression data: A survey. Knowledge and Data Engineering, IEEE Transactions
on, 16(11):1370–86. [cited at p. 129, 136]
[Jiang et al., 2009]
Jiang, Y., Xu, C., Dong, F., Yang, Y., Jiang, W., and Yang, S. (2009).
Disruption of the acetoacetate decarboxylase gene in solvent-producing clostridium
acetobutylicum increases the butanol ratio. Metabolic Engineering, 11(4-5):284 – 91.
[cited at p. 65, 78]
[Jones and Woods, 1986]
Jones, D. and Woods, D. (1986). Acetone-butanol fermentation
revisited. Microbiological Reviews, 50(4):484–524. [cited at p. 7]
[Jones et al., 2008]
Jones, S., Paredes, C., Tracy, B., Cheng, N., Sillers, R., Senger, R.,
and Papoutsakis, E. (2008). The transcriptional program underlying the physiology of
clostridial sporulation. Genome Biology, 9(7):R114. [cited at p. 21, 39, 43, 68, 93, 95]
[Joyce and Palsson, 2006]
Joyce, A. and Palsson, B. (2006). The model organism as a
system: integrating’omics’ data sets. Nature Reviews Molecular Cell Biology, 7(3):198–
210. [cited at p. 20, 31, 32, 77, 78, 117]
[Junghans, 2008]
Junghans, M. (2008). Visualization of hyperedges in fixed graph layouts.
Master’s thesis, Brandenburg University of Technology, Cottbus. [cited at p. 54]
[Junker et al., 2006]
Junker, B., Klukas, C., and Schreiber, F. (2006). Vanted: A system
for advanced data analysis and visualization in the context of biological networks. BMC
Bioinformatics, 7(1):109. [cited at p. 29]
[Junne, 2010]
Junne, S. (2010). Stimulus response experiments for modelling product
formation in Clostridium acetobutylicum fermentations. PhD thesis, Berlin Institute of
Technology, Germany, Berlin. [cited at p. 12, 13, 17, 80, 81, 82]
[Kansiz et al., 2001]
Kansiz, M., Gapes, J., McNaughton, D., Lendl, B., and Schuster,
K. (2001). Mid-infrared spectroscopy coupled to sequential injection analysis for the
on-line monitoring of the acetone-butanol fermentation process. Analytica Chimica
Acta, 438(1-2):175 – 86. [cited at p. 20]
BIBLIOGRAPHY 151
[Katsnelson, 2013]
Katsnelson, A. (2013). Momentum grows to make personalized medi-
cine more precise. Nature Medicine, 19(3):249. [cited at p. 78]
[Khatri et al., 2012]
Khatri, P., Sirota, M., and Butte, A. (2012). Ten years of path-
way analysis: Current approaches and outstanding challenges. PLoS Comput Biol,
8(2):e1002375. [cited at p. 34, 53, 77, 78]
[Kim et al., 1984]
Kim, B., Bellows, P., Datta, R., and Zeikus, J. (1984). Control of
carbon and electron flow in clostridium acetobutylicum fermentations: Utilization
of carbon monoxide to inhibit hydrogen production and to enhance butanol yields.
Applied and Environmental Microbiology, 48(4):764–70. [cited at p. 17]
[Klamt and Gilles, 2004]
Klamt, S. and Gilles, E. D. (2004). Minimal cut sets in bio-
chemical reaction networks. Bioinformatics, 20(2):226–234. [cited at p. 78]
[Klamt et al., 2007]
Klamt, S., Saez-Rodriguez, J., and Gilles, E. (2007). Structural and
functional analysis of cellular networks with cellnetanalyzer. BMC Systems Biology,
1(1):2. [cited at p. 31, 117]
[Klipp et al., 2007]
Klipp, E., Liebermeister, W., Helbig, A., Kowald, A., and Schaber, J.
(2007). Standards in computational systems biology. [cited at p. 3]
[Klipp et al., 2004]
Klipp, E., Liebermeister, W., and Wierling, C. (2004). Inferring dy-
namic properties of biochemical reaction networks from structural knowledge. Genome
Informatics Series, pages 125–37. [cited at p. 32, 78]
[Koepke et al., 2011]
Koepke, M., Mihalcea, C., Bromley, J., and Simpson, S. (2011).
Fermentative production of ethanol from carbon monoxide. Current Opinion in
Biotechnology, 22(3):320 – 5. [cited at p. 47]
[Koestler et al., 2010]
Koestler, T., von Haeseler, A., and Ebersberger, I. (2010). Fact:
Functional annotation transfer between proteins with similar feature architectures.
BMC Bioinformatics, 11(1):417. [cited at p. 56, 64]
[Kuit et al., 2012]
Kuit, W., Minton, N., Lopez-Contreras, A., and Eggink, G. (2012).
Disruption of the acetate kinase (ack) gene of clostridium acetobutylicum results in
delayed acetate production. Applied Microbiology and Biotechnology, 94(3):729–41.
[cited at p. 15, 20]
[Kumar et al., 2012] Kumar, A., Suthers, P., and Maranas, C. (2012). Metrxn: a know-
ledgebase of metabolites and reactions spanning metabolic models and databases. BMC
Bioinformatics, 13(1):6. [cited at p. 28, 38, 116]
[Lebedeva et al., 2012]
Lebedeva, G., Sorokin, A., Faratian, D., Mullen, P., Goltsov, A.,
Langdon, S., Harrison, D., and Goryanin, I. (2012). Model-based global sensitivity
analysis as applied to identification of anti-cancer drug targets and biomarkers of
drug resistance in the erbb2/3 network. European Journal of Pharmaceutical Sciences,
46(4):244 – 58. [cited at p. 106]
[Lee and Lee, 2009]
Lee, B. and Lee, D. (2009). Protein comparison at the domain
architecture level. BMC Bioinformatics, 10(Suppl 15):S5. [cited at p. 64, 71, 75]
152 BIBLIOGRAPHY
[Lee et al., 2012]
Lee, J., Jang, Y., Choi, S., Im, J., Song, H., Cho, J., Seung, D.,
Papoutsakis, E., Bennett, G., and Lee, S. (2012). Metabolic engineering of clostridium
acetobutylicum atcc 824 for isopropanol-butanol-ethanol fermentation. Applied and
Environmental Microbiology, 78(5):1416–23. [cited at p. 15]
[Lee et al., 2008a]
Lee, J., Yun, H., Feist, A., Palsson, B., and Lee, S. (2008a). Genome-
scale reconstruction and in silico analysis of the clostridium acetobutylicum atcc 824
metabolic network. Applied Microbiology and Biotechnology, 80:849–62.
[cited at p. 28,
36, 64, 80, 117]
[Lee et al., 2008b]
Lee, S., Park, J., Jang, S., Nielsen, L., Kim, J., and Jung, K. (2008b).
Fermentative butanol production by clostridia. Biotechnology and Bioengineering,
101(2):209–28. [cited at p. 6, 11, 14, 116]
[Lehmann et al., 2012a]
Lehmann, D., Hoenicke, D., Ehrenreich, A., Schmidt, M.,
Weuster-Botz, D., Bahl, H., and Luetke-Eversloh, T. (2012a). Modifying the product
pattern of clostridium acetobutylicum. Applied Microbiology and Biotechnology,
94(3):743–54. [cited at p. 15, 20, 56, 65, 78, 95, 103, 115]
[Lehmann and Luetke-Eversloh, 2011]
Lehmann, D. and Luetke-Eversloh, T. (2011).
Switching clostridium acetobutylicum to an ethanol producer by disruption of the butyr-
ate/butanol fermentative pathway. Metabolic Engineering, 13(5):464 – 73.
[cited at p. 13,
15, 21, 74, 115]
[Lehmann et al., 2012b]
Lehmann, D., Radomski, N., and Luetke-Eversloh, T. (2012b).
New insights into the butyric acid metabolism of clostridium acetobutylicum. Applied
Microbiology and Biotechnology, 96(5):1325–39. [cited at p. 15, 56, 65, 78, 95, 103]
[Lehmann et al., 1999]
Lehmann, T., Gonner, C., and Spitzer, K. (1999). Survey: Inter-
polation methods in medical image processing. Medical Imaging, IEEE Transactions
on, 18(11):1049–75. [cited at p. 88]
[Levandovski and D, 1971]
Levandovski, M. and D, W. (1971). Distance between sets.
Letters to Nature, 234:34–5. [cited at p. 58]
[Li et al., 2010]
Li, C., Donizelli, M., Rodriguez, N., Dharuri, H., Endler, L., Chelliah,
V., Li, L., He, E., Henry, A., Stefan, M., Snoep, J., Hucka, M., Le Novere, N., and
Laibe, C. (2010). BioModels Database: An enhanced, curated and annotated resource
for published quantitative kinetic models. BMC Systems Biology, 4:92. [cited at p. 3]
[Li et al., 2007]
Li, C., Yang, C., and Shan, H. (2007). Maximizing propylene yield by
two-stage riser catalytic cracking of heavy oil. Industrial & Engineering Chemistry
Research, 46(14):4914–20. [cited at p. 6]
[Liebermeister and Klipp, 2006]
Liebermeister, W. and Klipp, E. (2006). Bringing meta-
bolic networks to life: integration of kinetic, metabolic, and proteomic data. Theoretical
Biology and Medical Modelling, 3(1):42. [cited at p. 81]
[Lin et al., 2006]
Lin, K., Zhu, L., and Zhang, D. (2006). An initial strategy for comparing
proteins at the domain architecture level. Bioinformatics, 22(17):2081–6.
[cited at p. 56,
58, 74]
BIBLIOGRAPHY 153
[Luetke-Eversloh and Bahl, 2011]
Luetke-Eversloh, T. and Bahl, H. (2011). Metabolic en-
gineering of clostridium acetobutylicum: recent advances to improve butanol production.
Current Opinion in Biotechnology, 22(5):634 – 47. [cited at p. 10, 14]
[Lukashin and Fuchs, 2001]
Lukashin, A. and Fuchs, R. (2001). Analysis of temporal
gene expression profiles: clustering by simulated annealing and determining the optimal
number of clusters. Bioinformatics, 17(5):405–14. [cited at p. 129]
[Maddox et al., 2000]
Maddox, I., Steiner, E., Hirsch, S., Wessner, S., Gutierrez, N.,
Gapes, J., and Schuster, K. (2000). The cause of ”acid crash” and ”acidogenic
fermentations” during the batch acetone-butanol-ethanol(abe-) fermentation process.
Journal of Molecular Microbiology and Biotechnology, 2(1):95–100. [cited at p. 13]
[Mann and Luetke-Eversloh, 2013]
Mann, M. and Luetke-Eversloh, T. (2013). Thiolase
engineering for enhanced butanol production in clostridium acetobutylicum. Biotech-
nology and Bioengineering, 110(3):887–97. [cited at p. 14, 94, 95, 103, 117]
[Mariano et al., 2011]
Mariano, A., Qureshi, N., Filho, R., and Ezeji, T. (2011). Biopro-
duction of butanol in bioreactors: New insights from simultaneous in situ butanol
recovery to eliminate product toxicity. Biotechnology and Bioengineering, 108(8):1757–
65. [cited at p. 18]
[Marino et al., 2008]
Marino, S., Hogue, I., Ray, C., and Kirschner, D. (2008). A meth-
odology for performing global uncertainty and sensitivity analysis in systems biology.
Journal of Theoretical Biology, 254(1):178 – 96. [cited at p. 110]
[Mashego et al., 2007]
Mashego, M., Rumbold, K., Mey, M., Vandamme, E., Soetaert, W.,
and Heijnen, J. (2007). Microbial metabolomics: past, present and future methodologies.
Biotechnology Letters, 29(1):1–16. [cited at p. 20]
[Matthews et al., 2009]
Matthews, L., Gopinath, G., Gillespie, M., Caudy, M., Croft, D.,
de Bono, B., Garapati, P., Hemish, J., Hermjakob, H., Jassal, B., Kanapin, A., Lewis,
S., Mahajan, S., May, B., Schmidt, E., Vastrik, I., Wu, G., Birney, E., Stein, L., and
D’Eustachio, P. (2009). Reactome knowledgebase of human biological pathways and
processes. Nucleic Acids Research, 37(suppl 1):D619–D622. [cited at p. 28, 78]
[Meyer et al., 1986]
Meyer, C., Roos, J., and Papoutsakis, E. (1986). Carbon monoxide
gasing leads to alcohol production and butyrate uptake without acetone formation in
continuous cultures of clostridium acetobutylicum. Applied Microbiology and Biotech-
nology, 24(2):159–67. [cited at p. 13, 17]
[Meyer and Papoutsakis, 1989]
Meyer, C. L. and Papoutsakis, E. (1989). Continuous
and biomass recycle fermentations of clostridium acetobutylicum. Bioprocess and
Biosystems Engineering, 4:1–10. [cited at p. 19]
[Millat et al., 2013a]
Millat, T., Janssen, H., Bahl, H., Fischer, R., and Wolkenhauer, O.
(2013a). Integrative modelling of ph-dependent enzyme activity and transcriptomic
regulation of the acetonebutanolethanol fermentation of clostridium acetobutylicum in
continuous culture. Microbial Biotechnology, 6(5):526–39. [cited at p. 81]
154 BIBLIOGRAPHY
[Millat et al., 2013b]
Millat, T., Janssen, H., Thorn, G., King, J., Bahl, H., Fischer, R.,
and Wolkenhauer, O. (2013b). A shift in the dominant phenotype governs the ph-
induced metabolic switch of clostridium acetobutylicum in phosphate-limited continuous
cultures. Applied Microbiology and Biotechnology, 97(14):6451–66. [cited at p. 14]
[Monot et al., 1984]
Monot, F., Engasser, J., and Petitdemange, H. (1984). Influence
of ph and undissociated butyric acid on the production of acetone and butanol in
batch cultures of clostridium acetobutylicum. Applied Microbiology and Biotechnology,
19(6):422–6. [cited at p. 13]
[Monot et al., 1982]
Monot, F., Martin, J., Petitdemange, H., and Gay, R. (1982). Acet-
one and butanol production by clostridium acetobutylicum in a synthetic medium.
Applied and Environmental Microbiology, 44(6):1318–24. [cited at p. 16]
[Nacher et al., 2005]
Nacher, J., Yamada, T., Goto, S., Kanehisa, M., and Akutsu, T.
(2005). Two complementary representations of a scale-free network. Physica A:
Statistical Mechanics and its Applications, 349(1-2):349 – 63. [cited at p. 30]
[Nair et al., 1999]
Nair, R., Green, E., Watson, D., Bennett, G., and Papoutsakis, E.
(1999). Regulation of the sol locus genes for butanol and acetone formation in clostridium
acetobutylicumatcc 824 by a putative transcriptional repressor. Journal of Bacteriology,
181(1):319–30. [cited at p. 14]
[Nielsen et al., 2009]
Nielsen, D., Leonard, E., Yoon, S., Tseng, H., Yuan, C., and Prather,
K. (2009). Engineering alternative butanol production platforms in heterologous
bacteria. Metabolic Engineering, 11(4-5):262 – 73. [cited at p. 15]
[Nobeli and Thornton, 2006]
Nobeli, I. and Thornton, J. (2006). A bioinformatician’s
view of the metabolome. BioEssays, 28(5):534–45. [cited at p. 129]
[Noelling et al., 2001]
Noelling, J., Breton, G., Omelchenko, M., Makarova, K., Zeng, Q.,
Gibson, R., Lee, H., Dubois, J., Qiu, D., Hitti, J., Wolf, Y., Tatusov, R., Sabathe, F.,
Doucette-Stamm, L., Soucaille, P., Daly, M., Bennett, G., Koonin, E., and Smith, D.
(2001). Genome sequence and comparative analysis of the solvent-producing bacterium
clostridium acetobutylicum. Journal of Bacteriology, 183(16):4823–4838.
[cited at p. 10]
[Nolan et al., 2006]
Nolan, T., Hands, R., and Bustin, S. (2006). Quantification of mrna
using real-time rt-pcr. Nature protocols, 1(3):1559–82. [cited at p. 21]
[Oberhardt et al., 2009]
Oberhardt, M., Palsson, B., and Papin, J. A. (2009). Applic-
ations of genome-scale metabolic reconstructions. Molecular systems biology, 5(1).
[cited at p. 77]
[Ofran et al., 2005]
Ofran, Y., Punta, M., Schneider, R., and Rost, B. (2005). Beyond
annotation transfer by homology: novel protein-function prediction methods to assist
drug discovery. Drug Discovery Today, 10(21):1475 – 82. [cited at p. 56]
[¨
Ozilgen, 1988] ¨
Ozilgen, M. (1988). Kinetics of multiproduct acidogenic and solvent-
ogenic batch fermentations. Applied Microbiology and Biotechnology, 29(6):536–43.
[cited at p. 80]
BIBLIOGRAPHY 155
[Pagel et al., 2013]
Pagel, H., Ingwersen, J., Poll, C., Kandeler, E., and Streck, T. (2013).
Micro-scale modeling of pesticide degradation coupled to carbon turnover in the
detritusphere: model description and sensitivity analysis. Biogeochemistry, pages 1–20.
[cited at p. 106]
[Palsson and Zengler, 2010]
Palsson, B. and Zengler, K. (2010). The challenges of integ-
rating multi-omic data sets. Nature chemical biology, 6(11):787. [cited at p. 3]
[Papoutsakis et al., 1987]
Papoutsakis, E., Bussineau, C., Chu, I., Diwan, R., and Huese-
mann, M. (1987). Transport of substrates and metabolites and their effect on cell
metabolism (in butyric-acid and methylotrophic fermentations)a. Annals of the New
York Academy of Sciences, 506(1):24–50. [cited at p. 13, 65, 66]
[Papoutsakis, 1984]
Papoutsakis, E. T. (1984). Equations and calculations for ferment-
ations of butyric acid bacteria. Biotechnology and Bioengineering, 26(2):174–87.
[cited at p. 80, 116]
[Paredes et al., 2005]
Paredes, C. J., Alsaker, K. V., and Papoutsakis, E. T. (2005). A
comparative genomic view of clostridial sporulation and physiology. Nature Reviews
Microbiology, 3(12):969–78. [cited at p. 13]
[Patil and Nielsen, 2005]
Patil, K. and Nielsen, J. (2005). Uncovering transcriptional reg-
ulation of metabolism by using metabolic network topology. Proceedings of the National
Academy of Sciences of the United States of America, 102(8):2685–9.
[cited at p. 36, 78]
[Pattengale et al., 2010]
Pattengale, N., Alipour, M., Bininda-Emonds, O., Moret, B.,
and Stamatakis, A. (2010). How many bootstrap replicates are necessary? Journal of
Computational Biology, 17(3):337–54. [cited at p. 123]
[Peguin et al., 1994]
Peguin, S., Goma, G., Delorme, P., and Soucaille, P. (1994). Meta-
bolic flexibility of clostridium acetobutylicum in response to methyl viologen addition.
Applied Microbiology and Biotechnology, 42(4):611–6. [cited at p. 13]
[Peguin and Soucaille, 1995]
Peguin, S. and Soucaille, P. (1995). Modulation of Carbon
and Electron Flow in Clostridium acetobutylicum by Iron Limitation and Methyl
Viologen Addition. Appl. Environ. Microbiol., 61(1):403–5. [cited at p. 16]
[Pellegrini et al., 1999]
Pellegrini, M., Marcotte, E., Thompson, M., Eisenberg, D., and
Yeates, T. (1999). Assigning protein functions by comparative genome analysis: Protein
phylogenetic profiles. Proceedings of the National Academy of Sciences, 96(8):4285–8.
[cited at p. 64]
[Pfromm et al., 2010]
Pfromm, P., Amanor-Boadu, V., Nelson, R., Vadlani, P., and Madl,
R. (2010). Bio-butanol vs. bio-ethanol: A technical and economic assessment for
corn and switchgrass fermented by yeast or clostridium acetobutylicum. Biomass and
Bioenergy, 34(4):515 – 24. [cited at p. 7]
[Punta et al., 2012]
Punta, M., Coggill, P., Eberhardt, R., Mistry, J., Tate, J., Boursnell,
C., Pang, N., Forslund, K., Ceric, G., Clements, J., Heger, A., Holm, L., Sonnhammer,
E., Eddy, S., Bateman, A., and Finn, R. (2012). The pfam protein families database.
Nucleic Acids Research, 40(D1):D290–D301. [cited at p. 56]
156 BIBLIOGRAPHY
[Qureshi et al., 2006]
Qureshi, N., Li, X., Hughes, S., Saha, B., and Cotta, M. (2006).
Butanol production from corn fiber xylan using clostridium acetobutylicum. Biotech-
nology Progress, 22(3):673–80. [cited at p. 17]
[Qureshi and Maddox, 2005]
Qureshi, N. and Maddox, I. (2005). Reduction in butanol
inhibition by perstraction: Utilization of concentrated lactose/whey permeate by
clostridium acetobutylicum to enhance butanol fermentation economics. Food and
Bioproducts Processing, 83(1):43 – 52. [cited at p. 18]
[Rabitz et al., 1983]
Rabitz, H., Kramer, M., and Dacol, D. (1983). Sensitivity analysis in
chemical kinetics. Annual Review of Physical Chemistry, 34(1):419–461.
[cited at p. 106]
[Rao and Mutharasan, 1987]
Rao, G. and Mutharasan, R. (1987). Altered electron flow
in continuous cultures of clostridium acetobutylicum induced by viologen dyes. Applied
and Environmental Microbiology, 53(6):1232–35. [cited at p. 13]
[Ravagnani et al., 2000]
Ravagnani, A., Jennert, K., Steiner, E., Gruenberg, R., Jefferies,
J., Wilkinson, S., Young, D., Tidswell, E., Brown, D., Youngman, P., Morris, J., and
Young, M. (2000). Spo0a directly controls the switch from acid to solvent production
in solvent-forming clostridia. Molecular Microbiology, 37(5):1172–85. [cited at p. 14]
[Reardon and Bailey, 1989]
Reardon, K. and Bailey, J. (1989). Effects of ph and added
metabolites on bioconversions by immobilized non-growing clostridium acetobutylicum.
Biotechnology and Bioengineering, 34(6):825–37. [cited at p. 18]
[Reardon and Bailey, 1992]
Reardon, K. and Bailey, J. (1992). Activity regeneration
in continuous clostridium acetobutylicum bioconversions of glucose. Biotechnology
Progress, 8(4):316–26. [cited at p. 18]
[Reed et al., 2006]
Reed, J., Famili, I., Thiele, I., and Palsson, B. (2006). Towards
multidimensional genome annotation. Nature Reviews, 7:130–41. [cited at p. 34, 37, 38]
[Reuter, 2011]
Reuter, A. (2011). Datengetriebene forschung - herausforderung fr die
informatik. Spektrum der Wissenschaft - Extra, pages 6–9. [cited at p. 77]
[Rogers, 2002]
Rogers, P. (2002). Clostridia, Solvent Formation. John Wiley & Sons, Inc.
[cited at p. 10]
[Rost et al., 2003]
Rost, B., Liu, J., Nair, R., Wrzeszczynski, K., and Ofran, Y. (2003).
Automatic prediction of protein function. Cellular and Molecular Life Sciences CMLS,
60(12):2637–50. [cited at p. 64, 74]
[Saltelli et al., 2000]
Saltelli, A., Tarantola, S., and Campolongo, F. (2000). Sensitivity
analysis as an ingredient of modeling. Statistical Science, pages 377–95.
[cited at p. 106]
[Saltelli et al., 1999]
Saltelli, A., Tarantola, S., and Chan, K. (1999). A quantitative
model-independent method for global sensitivity analysis of model output. Technomet-
rics, 41(1):39–56. [cited at p. 110]
[Schaedel and Franco-Lara, 2009]
Schaedel, F. and Franco-Lara, E. (2009). Rapid
sampling devices for metabolic engineering applications. Applied Microbiology and
Biotechnology, 83(2):199–208. [cited at p. 20]
BIBLIOGRAPHY 157
[Schaffer et al., 2002]
Schaffer, S., Isci, N., Zickner, B., and Duerre, P. (2002). Changes
in protein synthesis and identification of proteins specifically induced during solvento-
genesis in clostridium acetobutylicum. Electrophoresis, 23(1):110–21. [cited at p. 21]
[Schaub, 2005]
Schaub, J. (2005). Isotopisch instation¨are 13C-Stoffflussanalyse in Es-
cherichia coli. PhD thesis, Universit¨at Stuttgart, Germany. [cited at p. 20, 104]
[Schaub et al., 2006]
Schaub, J., Schiesling, C., Reuss, M., and Dauner, M. (2006). Integ-
rated sampling procedure for metabolome analysis. Biotechnology Progress, 22(5):1434–
42. [cited at p. 20]
[Schmidt and Jirstrand, 2006]
Schmidt, H. and Jirstrand, M. (2006). Systems biology
toolbox for matlab: a computational platform for research in systems biology. Bioin-
formatics, 22(4):514–5. http://www.sbtoolbox2.org.[cited at p. 90]
[Schwarz et al., 2007a]
Schwarz, K., Fiedler, T., Fischer, R., and Bahl, H. (2007a). A
standard operating procedure (sop) for the preparation of intra- and extracellular
proteins of clostridium acetobutylicum for proteome analysis. Journal of Microbiological
Methods, 68(2):396 – 402. [cited at p. 21]
[Schwarz et al., 2012]
Schwarz, K., Kuit, W., Grimmler, C., Ehrenreich, A., and Kengen,
S. (2012). A transcriptional study of acidogenic chemostat cells of clostridium aceto-
butylicum - cellular behavior in adaptation to n-butanol. Journal of Biotechnology,
161(3):366 – 77. [cited at p. 17, 21]
[Schwarz et al., 2007b]
Schwarz, R., Liang, C., Kaleta, C., Kuhnel, M., Hoffmann, E.,
Kuznetsov, S., Hecker, M., Griffiths, G., Schuster, S., and Dandekar, T. (2007b).
Integrated network reconstruction, visualization and analysis using yanasquare. BMC
Bioinformatics, 8(1):313. [cited at p. 29, 117]
[SE, 2008] SE, B. (2008). Technisches merkblatt. Internet. M 2084 d. [cited at p. 6]
[Secrier and Schneider, 2013]
Secrier, M. and Schneider, R. (2013). Visualizing time-
related data in biology, a review. Briefings in Bioinformatics.[cited at p. 55]
[Senger and Papoutsakis, 2008]
Senger, R. and Papoutsakis, E. (2008). Genome-scale
model for clostridium acetobutylicum: Part i metabolic network resolution and analysis.
Biotechnology and Bioengineering, 101(5):1036–52. [cited at p. 28, 36, 80, 117]
[Shannon et al., 2003]
Shannon, P., Markiel, A., Ozier, O., Baliga, N., Wang, J., Ramage,
D., Amin, N., Schwikowski, B., and Ideker, T. (2003). Cytoscape: A software environ-
ment for integrated models of biomolecular interaction networks. Genome Research,
13(11):2498–504. [cited at p. 3, 31]
[Shinto et al., 2007]
Shinto, H., Tashiro, Y., Yamashita, M., Kobayashi, G., Sekiguchi, T.,
Hanai, T., Kuriya, Y., Okamoto, M., and Sonomoto, K. (2007). Kinetic modeling and
sensitivity analysis of acetone-butanol-ethanol production. Journal of Biotechnology,
131(1):45 – 56. [cited at p. 81, 88]
[Siegel and Himmele, 1980]
Siegel, H. and Himmele, W. (1980). Synthesis of intermediates
by rhodium-catalyzed hydroformylation. Angewandte Chemie International Edition in
English, 19(3):178–83. [cited at p. 6]
158 BIBLIOGRAPHY
[Sillers et al., 2009]
Sillers, R., Al-Hinai, M., and Papoutsakis, E. (2009). Aldehyde-
alcohol dehydrogenase and/or thiolase overexpression coupled with coa transferase
downregulation lead to higher alcohol titers and selectivity in clostridium acetobutyl-
icum fermentations. Biotechnology and Bioengineering, 102(1):38–49.
[cited at p. 14, 78,
102, 116]
[Sillers et al., 2008]
Sillers, R., Chow, A., Tracy, B., and Papoutsakis, E. (2008). Meta-
bolic engineering of the non-sporulating, non-solventogenic clostridium acetobutylicum
strain m5 to produce butanol without acetone demonstrate the robustness of the acid-
formation pathways and the importance of the electron balance. Metabolic Engineering,
10(6):321 – 32. [cited at p. 13, 15]
[Snoep and Olivier, 2002]
Snoep, J. and Olivier, B. (2002). Java web simulation (jws);
a web based database of kinetic models. Molecular Biology Reports, 29:259–63.
[cited at p. 3]
[Sobol, 2001]
Sobol, I. (2001). Global sensitivity indices for nonlinear mathematical
models and their monte carlo estimates. Mathematics and Computers in Simulation,
55(1-3):271 – 80. [cited at p. 110]
[Srere, 1987]
Srere, P. A. (1987). Complexes of sequential metabolic enzymes. Annual
Review of Biochemistry, 56(1):89–124. [cited at p. 64]
[Srivastava and Volesky, 1991]
Srivastava, A. and Volesky, B. (1991). Measurement and
regulation of the culture reduction state in clostridium acetobutylicum. Biotechnology
and Bioengineering, 38(2):181–90. [cited at p. 20]
[Stelling et al., 2002]
Stelling, J., Klamt, S., Bettenbrock, K., Schuster, S., and Gilles,
E. (2002). Metabolic network structure determines key aspects of functionality and
regulation. Letters to Nature, 420:190–3. [cited at p. 32]
[Sturn et al., 2002]
Sturn, A., Quackenbush, J., and Trajanoski, Z. (2002). Genesis:
cluster analysis of microarray data. Bioinformatics, 18(1):207–8. [cited at p. 68]
[Sumner et al., 2012]
Sumner, T., Shephard, E., and Bogle, I. (2012). A methodology
for global-sensitivity analysis of time-dependent outputs in systems biology modelling.
Journal of The Royal Society Interface, 9(74):2156–66. [cited at p. 120, 135]
[Tanya Rezler, 2012]
Tanya Rezler, M. L. (2012). Global n-butanol market to reach 4
mln tonnes by 2020. Internet. [cited at p. 6]
[Tavazoie et al., 1999]
Tavazoie, S., Hughes, J., Campbell, M., Cho, R., and Church, G.
(1999). Systematic determination of genetic network architecture. Nature genetics,
22(3):281–5. [cited at p. 64]
[Terracciano and Kashket, 1986]
Terracciano, J. and Kashket, E. (1986). Intracellular
conditions required for initiation of solvent production by clostridium acetobutylicum.
Applied and Environmental Microbiology, 52(1):86–91. [cited at p. 13]
[Terracciano et al., 1988]
Terracciano, J., Rapaport, E., and Kashket, E. (1988). Stress-
and growth phase-associated proteins of clostridium acetobutylicum. Applied and
Environmental Microbiology, 54(8):1989–95. [cited at p. 21]
BIBLIOGRAPHY 159
[Theocharidis et al., 2009]
Theocharidis, A., Dongen, S., Enright, A., and Freeman, T.
(2009). Network visualization and analysis of gene expression data using biolayout
express 3d. Nature Protocols, 4:1535–50. [cited at p. 31]
[Thormann et al., 2002]
Thormann, K., Feustel, L., Lorenz, K., Nakotte, S., and Duerre,
P. (2002). Control of butanol formation in clostridium acetobutylicum by transcriptional
activation. Journal of Bacteriology, 184(7):1966–73. [cited at p. 14]
[Tomas et al., 2004]
Tomas, C., Beamish, J., and Papoutsakis, E. (2004). Transcriptional
analysis of butanol stress and tolerance in clostridium acetobutylicum. Journal of
Bacteriology, 186(7):2006–18. [cited at p. 17]
[Tomas et al., 2003]
Tomas, C., Welker, N., and Papoutsakis, E. (2003). Overexpression
of groesl in clostridium acetobutylicum results in increased solvent production and
tolerance, prolonged metabolism, and changes in the cell’s transcriptional program.
Applied and Environmental Microbiology, 69(8):4951–65. [cited at p. 20]
[Troyanskaya, 2005]
Troyanskaya, O. (2005). Putting microarrays in a context: Integrated
analysis of diverse biological data. Briefings in Bioinformatics, 6(1):34–43.
[cited at p. 31,
38]
[Troyanskaya et al., 2001]
Troyanskaya, O., Cantor, M., Sherlock, G., Brown, P., Hastie,
T., Tibshirani, R., Botstein, D., and Altman, R. (2001). Missing value estimation
methods for dna microarrays. Bioinformatics, 17(6):520–5. [cited at p. 120]
[Tummala et al., 2003a]
Tummala, S., Junne, S., and Papoutsakis, E. (2003a). Antisense
rna downregulation of coenzyme a transferase combined with alcohol-aldehyde dehydro-
genase overexpression leads to predominantly alcohologenic clostridium acetobutylicum
fermentations. Journal of Bacteriology, 185(12):3644–53. [cited at p. 14, 20, 137]
[Tummala et al., 2003b]
Tummala, S., Junne, S., Paredes, C., and Papoutsakis, E.
(2003b). Transcriptional analysis of product-concentration driven changes in cel-
lular programs of recombinant clostridium acetobutylicumstrains. Biotechnology and
Bioengineering, 84(7):842–54. [cited at p. 21, 102, 117]
[Turk and Pentland, 1991]
Turk, M. and Pentland, A. (1991). Face recognition using
eigenfaces. In Computer Vision and Pattern Recognition, 1991. Proceedings CVPR’91.,
IEEE Computer Society Conference on, pages 586–91. IEEE. [cited at p. 135]
[van Dongen, 2000]
van Dongen, S. (2000). Graph Clustering by Flow Simulation. PhD
thesis, University of Utrecht, Netherlands. [cited at p. 31]
[van Riel, 2006]
van Riel, N. A. (2006). Dynamic modelling and analysis of biochem-
ical networks: mechanism-based models and model-based experiments. Briefings in
Bioinformatics, 7(4):364–374. [cited at p. 77, 106, 117, 121]
[Vasconcelos et al., 1994]
Vasconcelos, I., Girbal, L., and Soucaille, P. (1994). Regulation
of carbon and electron flow in clostridium acetobutylicum grown in chemostat culture at
neutral ph on mixtures of glucose and glycerol. Journal of Bacteriology, 176(5):1443–50.
[cited at p. 95]
160 BIBLIOGRAPHY
[Vaz and Vicente, 2007]
Vaz, A. and Vicente, L. (2007). A particle swarm pattern search
method for bound constrained global optimization. Journal of Global Optimization,
39:197–219. 10.1007/s10898-007-9133-5. [cited at p. 94]
[Vijesh et al., 2013]
Vijesh, N., Chakrabarti, S., and Sreekumar, J. (2013). Modeling of
gene regulatory networks: A review. Journal of Biomedical Science and Engineering,
6:223 – 31. [cited at p. 117]
[Votruba et al., 1986]
Votruba, J., Volesky, B., and Yerushalmi, L. (1986). Mathematical
model of a batch acetone-butanol fermentation. Biotechnology and Bioengineering,
28(2):247–55. [cited at p. 80]
[W and BM, 2007]
W, B. and BM, P. (2007). Personalized medicine in the era of genomics.
JAMA, 298(14):1682–1684. [cited at p. 78]
[Waltemath et al., 2011]
Waltemath, D., Adams, R., Bergmann, F., Hucka, M., Kolpakov,
F., Miller, A., Moraru, I., Nickerson, D., Sahle, S., Snoep, J., and Le Novere, N.
(2011). Reproducible computational biology experiments with sed-ml - the simulation
experiment description markup language. BMC Systems Biology, 5(1):198.
[cited at p. 3,
118]
[Wang et al., 2011]
Wang, S., Zhang, Y., Dong, H., Mao, S., Zhu, Y., Wang, R., Luan,
G., and Li, Y. (2011). Formic acid triggers the ”acid crash” of acetone-butanol-ethanol
fermentation by clostridium acetobutylicum. Applied and Environmental Microbiology,
77(5):1674–80. [cited at p. 13]
[Wiesenborn et al., 1989]
Wiesenborn, D., Rudolph, F., and Papoutsakis, E. (1989).
Phosphotransbutyrylase from clostridium acetobutylicum atcc 824 and its role in
acidogenesis. Applied and Environmental Microbiology, 55(2):317–22. [cited at p. 12]
[Wrzodek et al., 2013]
Wrzodek, C., Buchel, F., Ruff, M., Drager, A., and Zell, A. (2013).
Precise generation of systems biology models from kegg pathways. BMC Systems
Biology, 7(1):15. [cited at p. 29]
[Yang et al., 2003]
Yang, H., Haddad, H., Tomas, C., Alsaker, K., and Papoutsakis, T.
(2003). A segmental nearest neighbor normalization and gene identification method
gives superior results for dna-array analysis. Proceedings of the National Academy of
Sciences, 100(3):1122–7. [cited at p. 37, 53]
[Yang et al., 2005]
Yang, Y., Xiao, Y., and Segal, M. (2005). Identifying differentially
expressed genes from microarray experiments via statistic synthesis. Bioinformatics,
21(7):1084–93. [cited at p. 78]
[Yerushalmi et al., 1983]
Yerushalmi, L., Volesky, B., Leung, W., and Neufeld, R. (1983).
Variations of solvent yield in acetone-butanol fermentation. European journal of applied
microbiology and biotechnology, 18(5):279–86. [cited at p. 80]
[Yerushalmi et al., 1986a]
Yerushalmi, L., Volesky, B., and Votruba, J. (1986a). Modelling
of culture kinetics and physiology for c. acetobutylicum. The Canadian Journal of
Chemical Engineering, 64(4):607–16. [cited at p. 80]
BIBLIOGRAPHY 161
[Yerushalmi et al., 1986b]
Yerushalmi, L., Volesky, B., and Votruba, J. (1986b). Systems
analysis of the culture physiology in acetone-butanol fermentation. Biotechnology and
Bioengineering, 28(9):1334–47. [cited at p. 11, 80]
[Yerushalmi et al., 1988]
Yerushalmi, L., Volesky, B., and Votruba, J. (1988). Fermenta-
tion process diagnosis using a mathematical model. Applied Microbiology and Biotech-
nology, 29(2-3):186–97. [cited at p. 80]
[Yeung et al., 2001]
Yeung, K., Haynor, D., and Ruzzo, W. (2001). Validating clustering
for gene expression data. Bioinformatics, 17(4):309–18. [cited at p. 71]
[Yeung and Ruzzo, 2001]
Yeung, K. and Ruzzo, W. (2001). Principal component analysis
for clustering gene expression data. Bioinformatics, 17(9):763–74.
[cited at p. 120, 130,
133, 136]
[Yeung et al., 2002]
Yeung, M., Tegn´er, J., and Collins, J. (2002). Reverse engineering
gene networks using singular value decomposition and robust regression. Proceedings
of the National Academy of Sciences, 99(9):6163–8. [cited at p. 120]
[You and Yin, 2000]
You, L. and Yin, J. (2000). Patterns of regulation from mrna and
protein time series. Metabolic engineering, 2(3):210–7. [cited at p. 129]
[Zhao et al., 2005]
Zhao, Y., Tomas, C. A., Rudolph, F. B., Papoutsakis, E., and Bennett,
G. N. (2005). Intracellular butyryl phosphate and acetyl phosphate concentrations in
clostridium acetobutylicum and their implications for solvent formation. Applied and
Environmental Microbiology, 71(1):530–7. [cited at p. 13, 95]
[Zheng and Rundell, 2006]
Zheng, Y. and Rundell, A. (2006). Comparative study of
parameter sensitivity analyses of the tcr-activated erk-mapk signalling pathway. IEE
Proceedings-Systems Biology, 153(4):201–11. [cited at p. 110]
[Zi et al., 2005]
Zi, Z., Cho, K., Sung, M., Xia, X., Zheng, J., and Sun, Z. (2005). In
silico identification of the key components and steps in ifn-
γ
induced jak-stat signaling
pathway. FEBS Letters, 579(5):1101 – 8. [cited at p. 106]
Appendices
163
Appendix A
Dynamic Model Equations
dce
Glc
dt =−r0|1
dci
ACoA
dt =ρ
cX
(r0|1Yp,Glc −r1|11 +r11|1+r2,3|1,7−r1|3−r1|9)−µ·ci
ACoA
dce
ACE
dt =r11|2−r2|11 −r2,3|1,7−D·ce
ACE
dci
AACoA
dt =ρ
cX
(0.5r1|3−r3|7−r3|4)−µ·ci
AACoA
dci
BCoA
dt =ρ
cX
(r3|4+r6,3|4,7−r4|5+r5|4−r4|10)−µ·ci
BCoA
dci
BUP
dt =ρ
cX
(−r5|4+r4|5+r6|5−r5|6)−µ·ci
BUP
dce
BU
dt =r5|6−r6|5−r6,3|4,7−D·ce
BU
dci
AA
dt =ρ
cX
(r3|7−r7|8)−µ·ci
AA
dce
ACN
dt =r7|8−D·ce
ACN
dce
ETOH
dt =r1|9−D·ce
ETOH
dce
BUOH
dt =r4|10 −D·ce
BUOH
dci
ACP
dt =ρ
cX
(r1|11 −r11|1)−µ·ci
ACP
165
166 APPENDIX A. DYNAMIC MODEL EQUATIONS
r1|11 =f(C1742)·cX·k1|11
ci
ACoA
Km1|11 +ci
ACoA
r11|1=f(C1742)·cX·k11|1
ci
ACP
Km1|11 +ci
ACP
r11|2=f(C1743)·cX·k11|2
ci
ACP
Km11|2+ci
ACP
r2|11 =f(C1743)·cX·k2|11
ce
ACE
Km2|11 +ce
ACE
r2,3|1,7=f(P0163,0164)·cX·k2,3|1,7
ce
ACE
Km2,3|1,7+ce
ACE
ci
AACoA
Km2,3|1,7+ci
AACoA
r1|3=f(C2873)·cX·k1|3
ci
ACoA
2
Km1|32+ Kn1|3ci
ACoA + (ci
ACoA)2
r3|4=f(C2708,2711,2712)·cX·k3|4
ci
AACoA
Km3|4+ci
AACoA +(ci
AACoA)2
Ki3|4
r4|5=f(C3076)·cX·k4|5
ci
BCoA
Km4|5+ci
BCoA +ci
BUP
Ki4|5
r5|4=f(C3076)·cX·k5|4
ci
BUP
Km5|4+ci
BUP
r5|6=f(C3075)·cX·k5|6
ci
BUP
Km5|6+ci
BUP
r6|5=f(C3075)·cX·k6|5
ce
BU
Km6|5+ce
BU
r7|8=f(P0165)·cX·k7|8
ci
AA
Km7|8+ci
AA
r1|9=f(C3298,3299,P0162)·cX·k1|9
ci
ACoA
Km1|9+ci
ACoA
r4|10 =f(C3298,3299,P0162)·cX·k4|10
ci
BCoA
Km4|10 +ci
BCoA
Ki10|4,10
Ki10|4,10 +ce
BUOH
r3|10 =r2,3|1,7+r6,3|4,7
r6,3|4,7=f(CAP0163,P 0164)·cX·kce
BU ·ci
AACoA
Km + ci
AACoA ·ce
BU + KnAci
AACoA ·(1 + ci
AACoA
KiA ) + KnBce
BU ·(1 + ce
BU
KiB )
Appendix B
Scripts
B.1 Taverna Workflows
In this section the design and operation of work-flows with Taverna [
Hull et al., 2006
]
is introduced.
Taverna is a free work-flow management system based on Java. It offers access
to various web-services by third-parties through the Web Service Description
Language (WSDL). Such parties are e.g. the European-Bioinformatics Institute
from the European Molecular Biology Laboratory (EMBL-EBI), the National
Center for Biotechnology Information (NCBI), Kyoto Encyclopedia of Genes
and Genomes (KEGG) and BioMart. Furthermore, local services, like Beanshell
scripts, Java API, R scripts and Excel interaction, support the processing of
automated database queries. A list of accessible services is given on biocatalogue.
Taverna Workflow Services
Access to web services and scripts is granted via Simple Object Access Protocols
(SOAPs) [
?
] or Representational State Transfer (REST) services [
?
]. Both consist
of a defined number of input ports, the processing of inputs and the output ports
for the query results. A sequence of services is connected by linking the output
and input ports and, if necessary, by introducing formatting services.
Obtaining The Reactome
Since the recent change of accessibility of the KEGG-API, the here treated work-
flows needed entire re-structuring. Luckily, things got easier:
Download of the organism-specific maps from genes to enzymes, to reactions,
to compounds are possible from the respective sites through copy paste. From
these three lists the unique reaction identifiers serve as input for the reaction-
167
168 APPENDIX B. SCRIPTS
pair mapping workflow. It extracts the reactions with a REST-service first, then
it extracts the different RPAIR-identifiers for each reaction-identifier. Similar to
this workflow, annotation of compounds and genes is performed (annotation genes,
annotation compounds).
Phylogenetic Comparison
Given a specific enzyme number, the gene-identifiers for all annotated organism
are retrieved in KEGG, this is done by the enzyme in organism-workflow. Due
to formatting issues, the list of genes from this workflow is reorganised by a
MATLAB-script, enzymelistconversion.m, before Cytoscape can be used.
B.2. STATIC MODEL SCRIPTS 169
B.2 Static Model Scripts
B.2.1 Model Creation
The determineboundaryparameter.m script serves for the data generation to
determine the boundary parameter
b
that is used for integration of data to the
reactome graph. It requires five arguments at maximum:
•
the reactome database from KEGG, with reaction-IDs in the first, gene-IDs
in the second, and reaction pairs (cpd: cpd:) in the third column,
•
a data matrix, with genes in the rows and experimental conditions in the
columns,
•
an optional string argument
H
-graph to evaluate the boundary parameter
for the augmentation, otherwise the G-graph is calculated,
•
an optional column-number which data column corresponds to state
s1
,
otherwise it is the first,
•
an optional column-number which data column corresponds to state
s2
,
otherwise it is the last column.
The output is of this script a three dimensional data-cube, the two variable
dimensions are values of
b
and time of the data. From these the number of active
genes, the number of active reactions, the number of active metabolites, and the
edges to nodes fraction are calculated.
figure show determination b.m then provides a visualisation of the data-cube.
CreateFilteredGraphs.m
Integration of data into the reactome database is done by this script, it requires
seven input parameters:
1. the output directory,
2. a data matrix as before,
3. the reactome database as before,
4. the column-number for the first state,
5. the column-number for the second state,
6. boundary parameter b,
7. a string, assessing how to augment:
170 APPENDIX B. SCRIPTS
•both, both states serve for augmentation
•simple, no augmentation
•after, the first state only serves for augmentation
•before, the second state only serves for augmentation.
The output creates several nodelists that are converted into graphs by Cyto-
scape: the graph in states
s1
,
s2
, the graph for all reaction occurring in neither
state, and the graphs for all reactions activated or inactivated in both states.
B.2.2 Creation of a Comparison Database
Creation of the comparison database from chapter 3.4.1 requires the following lists
from two organisms: a list of genes, the annotation databases as downloaded from
taverna, and the mappings of genes to reactions. From these three, the Cytoscape
maps are constructed. An example file is given in bsu cac comp.m.
B.2.3 Creation of 3-HBDH Candidates
This analysis requires one approach to compare pfam-motifs: A script to generate
a map of pfam-motifs from the gene-annotation (CreateGeneMotifMAP.m). This
map is suited for import into Cytoscape. A script counts the occurrences of genes
and motifs in the map (CountOccurrences.m). Selection of the most frequent motifs
or genes is carried out by FindFrequentMotifs.m. For the clustering it is necessary
to find genes in the same clusters over different data (IdentifySameClusterGenes.m)
and a method to create subsets of the reactome in terms of the candidate genes
and their pfam-motifs (SelectFromGeneMotifMap.m). Finally the conversion into
matrix-format allows saving the results (SaveMapAsMatrix).
Inputs for these methods are simple lists of genes that are derived from the
analyses in other programmes.
The example script Search 3HBDH.m shows the automated candidate generation
with the help of the beforehand mentioned scripts.
B.3. DYNAMIC MODEL SCRIPTS 171
B.3 Dynamic Model Scripts
Model creation requires two computational steps, the conversion of the Reactome
database into the SBTOOLBOX2-format (ConvertKEGG2SBToolbox2.m) and
the supplying of the model with data (ConvertStdModel2SBT.m).
B.3.1 Converting The KEGG-Database Into The Standard Format
For conversion from KEGG into the standard format in the KEGG2SBTOOLBOX2 -
script three data-inputs are required:
•
the local database as discussed earlier that contains the mapping of tran-
scripts to reactions
•
the model file that contains all desired reactions and compounds, this can
be achieved either manually or by using Cytoscape.
•a list of extracellular compounds
The deposited script allows the computation of a complete standard model with
all transcripts at place, Michaelis-Menten type kinetics for multiple substrates,
if necessary and standardised parameters for the reactions, following the same
scheme as for reaction-identifiers. It is not possible to manually create reactions
that are not present in the database, if such a reaction is desired, it has to be in
both files, the database and the model.
B.3.2 Integrating Data Into The Standard Format
Integration is performed by the ConvertStdModel2SBT -script. It requires four
files
1. the standard-model file,
2. the transcriptome level timeseries data,
3. the glucose consumption profile,
4. the optical density profile
and a parameter that controls the integration of transcriptome level data that
either is implemented directly or via simplified via PCA reduction (5.1).
B.3.3 Model Simulation
Parameter estimation of the models is readily done by SBPD and the SBPD
file structure. For model comparison, several data were fed into the constructed
models. For model validation, several parameter variations were performed.
172 APPENDIX B. SCRIPTS
B.3.4 Sensitivity Analysis
Local Sensitivity Analysis
The local sensitivity analyses requires the symbolic computing package from
MATLAB. The DeriveModel.m file takes a SBTOOLBOX-model and calculates
the derivatives of the maximal velocities, integrates them back into the model to
allow simulation.
Global Sensitivity Analysis
The global sensitivity analysis is readily available through SBTOOLBOX-scripts.
In order to allow temporally resolved indices, a calculation script (MySensitiv-
ityAnalysis2.m) was wrapped around this script to allow the analysis of time
intervals and the SBTOOLBOX-scripts (SBsensglobalfast.m,rel sensglobaldefault-
objectiveSB.m) were adapted. This adaption includes the calculation of the
FAST-alternations for the whole time interval once. Then the calculation is split
into intervals and the sensitivity indices calculated per interval.
B.4. PRINCIPAL COMPONENT ANALYSIS 173
B.4 Principal Component Analysis
B.4.1 Dynamic Features
The dynamic features extraction and calculation is performed by the dy-
namic features.m-script. It is used to change three dynamical features by adding
150% or substracting 50% of the original parameter
B.4.2 Clustering
master script.m
This is the main calling script of the clustering algorithm, all parameters are
supplied in this module, the input and output directories, the data input and
the genes input (in case not the whole data set is to be used). Then the data
filter options, nonan,nT, how many missing values are admissible in a single
transcript expression profile and how many time points are to be taken. Then the
PCA-parameters, ncoeff,ssec and lsec, how many components are to be considered
and how the PC-space
C
should be partitioned. As these two numbers are the
numbers of partition in one half-space, they are dissected into 2
ssec ≤
2
lsec
parts.
Finally, the percentage of relevant data used to do the PCA is another input.
Filtering will be applied on two levels, the maximal span of the data and the
information content. Last but not least, the name of the annotation datafile is
supplied.
myGetGenData.m
This file imports the data matrix, with gene-identifiers in the rows, and the
temporal dimensions in the columns. A struct is handled back, it corresponds to
the specific filtering options with maximal nonan missing values of length nT.
mySaveGenData.m
This routine saves the struct obtained from myGetGenData.m for the imputation
or from the file obtained from the imputation back into a csv-file.
Imputation
Imputation is performed based on a web-service offered by the MPI-Potsdam
called MetaGeneAlyse [
Daub et al., 2003
]. The output-file from mySaveGenData
can be readily uploaded after manual replacement of occurring ”NaN” into ”NA”.
This website offers the opportunity to impute missing points according to three
different analyses based on Principal component analysis.
174 APPENDIX B. SCRIPTS
MyDataReduction.m
For meaningful data reduction we use a percentile approach. The lowest dynamics
and the highest entropy content are filtered out from the respective distributions
at a level of perc. These scripts are achieved using the Statistics Toolbox by
MATLAB. Filtered out genes are written out in individual files.
myPCA2.m
This script is the core module where the evaluation of the data’s properties hap-
pen.
It starts with the mapping of transcript expression profiles into the principal
component space of size ncoeff. For small ncoeff, one can generate a typographic
visualisation of different combinations of principal components, facilitating the
overview, what type of dynamic behaviours can be achieved.
In the next script, mycalculus.m, the individual transcript expression profile coef-
ficients are mapped on a sphere and the angles with respect to the first coefficient
are calculated. Subsequently, these angles are grouped into the pre-defined sectors
that represent the partition of the sphere into 2
ssec
large parts up to 2
lsec
small
parts.
Transcript profiles with similar dynamics are grouped using the script Simil-
arGenes.m which compares the vectors of sector membership with each other and
dependent and for rco=0 angular traits are calculated. For rco¿0 neighbouring
sectors are taken into account.
For transcript profiles of inverse dynamics, the same script compares the vector
of sector membership to the reflected sector.
These information are then evaluated in the script SharedClusters.m.
Clustering and Results
Finally, for any parameter rco, the program expects a cluster table file from
BioLayoutExpress 3D, that is generated by using the MCL-algorithm. This file is
then used to generate maps of cluster-identifiers to gene-identifiers to the data.
