Available online at www.sciencedirect.

com

ScienceDirect
IFAC PapersOnLine 55-20 (2022) 641–646
Genome-scale Metabolic Model Guided Subtyping Lung Cancer towards
Genome-scale
Genome-scale Metabolic
Metabolic Model
Model Guided
Guided Subtyping
Subtyping Lung
Lung Cancer
Cancer towards
towards
Genome-scale
Genome-scale Metabolic
Metabolic Personalized
Model
Model Guided
Guided Diagnosis
Subtyping
Subtyping Lung
Lung Cancer
Cancer towards
towards
Genome-scale Metabolic Personalized
Model Guided
Personalized Diagnosis
Subtyping
Diagnosis Lung Cancer towards
Personalized
Ezgi Personalized
Diagnosis
Diagnosis
Personalized Diagnosis
Tanıl, Nehir Kızılilsoley,
Ezgi
Emrah Nikerel*
Ezgi Tanıl,
Tanıl, Nehir
Nehir Kızılilsoley,

Kızılilsoley, Emrah
Emrah Nikerel*
Nikerel*
Ezgi Tanıl, 
Nehir Kızılilsoley,
Kızılilsoley, Emrah Nikerel*
Nikerel*
*Department of Genetics Ezgi Tanıl,
Ezgi Tanıl, Nehir
and Nehir 
Bioengineering,
Kızılilsoley, Emrah
Yeditepe
Emrah University,
Nikerel* İstanbul, Turkey

*Department
*Department (Tel:of Genetics
of+90216
Genetics and
and Bioengineering,
5780619; e-mail: Yeditepe
Yeditepe University,
University, İstanbul,
 emrah.nikerel@yeditepe.edu.tr).
Bioengineering, İstanbul, Turkey Turkey
*Department
*Department (Tel: of
of Genetics
+90216
Genetics and
5780619;
and Bioengineering,
e-mail:
Bioengineering, Yeditepe University,
emrah.nikerel@yeditepe.edu.tr).
Yeditepe University, İstanbul,
İstanbul, Turkey
Turkey
*Department (Tel:of+90216
Genetics 5780619; e-mail: emrah.nikerel@yeditepe.edu.tr).
and Bioengineering, Yeditepe University, İstanbul, Turkey
(Tel:
(Tel: +90216 5780619; e-mail: emrah.nikerel@yeditepe.edu.tr).
(Tel: +90216
+90216 5780619;
5780619; e-mail: e-mail: emrah.nikerel@yeditepe.edu.tr).
emrah.nikerel@yeditepe.edu.tr).
Abstract: Mathematical modeling of biological systems are useful in (i) gaining better understanding
Abstract:
about
Abstract: Mathematical
the physiological
Mathematicalprocesses modeling
modelingin of an biological
of organism, (ii)
biological systems
systems are
are useful
simulating in
in (i)
alternative
useful gaining
gaining better
(i) scenarios, better understanding
(iii) finding targets
understanding
Abstract:
about
for the
improved
Abstract: Mathematical
physiological
Mathematical
performance modeling
processes
modeling
within in of
an
of biological
organism,
biological
metabolic systems
(ii)
systems
engineering are
simulating
are useful
useful
context in
alternative
in
(iv) (i)
(i) gaining
scenarios,
gaining
performing better
better(iii) understanding
finding
several targets
understanding
functional
about the physiological
Abstract: Mathematical processes
modeling in of
an biological
organism, (ii) systemssimulating
are alternative
useful in (i) scenarios,
gaining better(iii) finding targets
understanding
about
for
about
analyses, the
improved
the physiological
performance
physiological
e.g. identify processes
within
processes
drug targets in
in an
an
(v) organism,
metabolic
organism,
process (ii)
engineering
(ii)
scheduling simulating
simulating context
within alternative
(iv)
alternative
the context scenarios,
performing
scenarios,
of industrial (iii)
(iii) finding
several
finding targets
functional
targets
biotechnology
for improved
about the performance
physiological within
processes in metabolic
an organism, engineering
(ii) simulating context (iv) performing
alternative scenarios, several
(iii) functional
finding targets
for
for
etc. improved
analyses,
improved e.g.
Increasing performance
identify
performance
the drug
predictive within
targets
within
capabilitymetabolic
(v) process
metabolic
of engineering
scheduling
engineering
these models is context
within
context
of the
common (iv)
context
(iv) performing
of
performing
interest industrial
within several
several
systems functional
biotechnology
functional
biology
analyses,
for improved e.g. identify
performance drug targets
within (v) process scheduling
metabolic engineering within
context the (iv)
context of industrial
performing biotechnology
several functional
analyses,
etc. e.g.
Increasing identify
the drug targets (v) process scheduling within the context of industrial biotechnology
analyses,
studies
etc. e.g.
Increasing
analyses, the predictive
e.g. identify
which allows
identify drug
drug targets
identification
predictive capability
targets (v)
of more
capability
(v)
of
of these
process
process
models
scheduling
effective
these models
scheduling and is of
of common
iswithin
personalized
within the
the context
common interest
treatmentof
interest
context
within
of industrial
strategies
within
industrial
systems biology
biotechnology
for complex
systems biology
biotechnology
etc.
studies
etc. Increasing
which
Increasing
metabolic diseasesthe
allows
the predictive
identification
predictive
such as capability
of
capability
cancer by of
more
of these
these
investigation models
effective
models
of and
diseaseis
is of
of common
personalized
common
metabolism interest
treatment
interest
and within
strategies
within
providing systems
for
systems
correct biology
complex
biology
subtyping
studies
etc. which
Increasing allows identification
the predictive of
capability more effective
of these models and personalized
ispersonalized
of commontreatment treatment strategies
interest within for
systems complex
biology
studies
metabolic
studies
and which
which
staging. By allows
diseases
allowssuchidentification
as cancer
identification
transforming by
gene-levelof
of more
more effective
investigation
effective
information of toand
disease
and metabolism
personalized
flux/metabolite and
treatment
level strategies
providing
strategies
information, correctfor
for
current complex
subtyping
complex
disease
metabolic
studies diseases
which allowssuch as cancer
identification by ofinvestigation
more effective of disease
and metabolism
personalized and
treatmentproviding
strategies correctfor subtyping
complex
metabolic
and
state staging.
metabolic
can bediseases
By
diseases
analyzed such
transforming
suchand asdiagnosis
as cancer
cancer byofinvestigation
gene-level
by investigation
information
cancer subtype of to
of disease
disease
can be metabolismlevel
flux/metabolite
metabolism
performed and information,
and
using providing
providing
a less invasive correct
current
correct subtyping
disease
subtyping
methods.
and staging.
metabolic By
diseases transforming
such as cancer gene-level information
by investigation to
of to flux/metabolite
disease metabolismlevel level information,
and information,
providing correct current disease
subtyping
and
state
and staging.
can
staging.be By
By transforming
analyzed and
transforming gene-level
diagnosis of
gene-level information
cancer subtype
information can
to flux/metabolite
be performed
flux/metabolite using
level a less invasive
information, current disease
methods.
current disease
statestaging.
and can be analyzed
By and diagnosis
transforming of cancer
gene-level subtypeto
information can be performed level
flux/metabolite usinginformation,
a less invasive methods.
current disease
state
In this
state can
can be
study,
be analyzed
subtyping
analyzed and
and diagnosis
and staging
diagnosis of
of cancer
of subtype
lungsubtype
cancer cancer,can can be performed
thatbeisperformed
one of the using
usingmain a
a less
causes
less invasive
invasive methods.
of cancer
methods.related
state
In can be analyzed and diagnosis of cancer subtype can be performed using a less invasive methods.
In this
this study,
deaths, subtyping
was performed
study, subtypingby and staging
staging of
andintegrating lung
lung cancer,
ofpublicly available
cancer, that is
is one
that RNAseqone of the
ofdata
the of main
main causes
normal,
causeslung of
of cancer
adenoma
cancer related
and
related
In
In this
deaths,
this study,
was
study,
adenocarcinomas subtyping
performed
subtyping
and lungand
by
and staging
integrating
staging
squamous of
of lung
publicly
lung
cell cancer,
available
cancer,
neoplasms that
that
to is
is one
RNAseq
one
human of
of the
data
the
genome main
of
mainscalecauses
normal,
causes
metabolic of
lung
of cancer
adenoma
cancer
model related
and
related
and
deaths,
In was performed
this study, subtypingby andintegrating
staging ofpublicly available
lung cancer, that RNAseq
is one ofdata the of mainnormal,
causeslung adenoma
of cancer related
deaths,
deaths, was
adenocarcinomas
was
classification performed
and
performed
of obtained by
lung
by integrating
squamous
integrating
flux distributions publicly
cell
publiclyusingavailable
neoplasms
available
linear to RNAseq
human
RNAseq
support data
genome
data
vector of
of normal,
scale
normal,
machine (SVM) lung
metabolic
lung adenoma
model
adenoma and
classifications.
adenocarcinomas
deaths, was performed and lung squamous publicly
by integrating cell neoplasms
available to RNAseq
human genome data of scale normal, metabolic
lung adenoma model and
adenocarcinomas
classification of and lung squamous cell neoplasms to human genome scale metabolic model and
adenocarcinomas
Differential flux
classification
adenocarcinomas of obtained
and
analysis
obtained
and lung
lung
flux
and
flux distributions
squamous
pathway
distributions
squamous cell
cell
using
neoplasms
enrichment
using
neoplasms
linear
methods
linear support
to human
support
to showed
human
vector
vector thatmachine
genome
genome scale
model
machine
scale
(SVM)
metabolic
adequately
(SVM)
metabolic
classifications.
model
model and
represented
classifications.
and
classification
Differential
classification
tumour metabolism. of
flux
of obtained
analysis
obtained
SVMand flux
and
flux distributions
pathway
distributions
classification using
enrichment
using
accuracies linear
methods
linear support
weresupport showed
calculated vector
vector that
as machine
model
machine (SVM)
adequately
(SVM) classifications.
represented
classifications.
Differential
classification flux
of analysis
obtained flux pathway
distributions enrichment
using methods
linear supportshowed vector thatmore
model
machine thanadequately
99% for
(SVM) normal
representedand
classifications.
Differential
tumour
Differential
cancer cells flux
metabolism.
flux
and analysis
SVM
analysis
94% for and
and
adenomaspathway
classification
pathway and enrichment
accuracies
enrichment
adenocarcinomas methods
were
methods and showed
calculated
showed
squamous that
as
that model
more
model
cell thanadequately
99%
adequately
neoplasms, for represented
normal
represented
indicating and
high
tumour
Differential metabolism.
flux analysisSVM classification
and accuracies
pathway enrichment were
methods calculated
showed as as more
thatmore
model than 99%
adequately for normal
representedand
tumour
cancer metabolism.
cells and SVM classification accuracies were calculated than 99% for normal and
tumour
predictive
cancer
tumour cells and 94%
metabolism.
capability
metabolism. 94%of for
SVM
SVM flux
for adenomas and
and adenocarcinomas
classification
distributions.
adenomas
classification accuracies
accuracies were
adenocarcinomas and
and squamous
were calculatedsquamous
calculated as
as more
cell
cell neoplasms,
more than
than 99%99% for
neoplasms, indicating
normal
indicating
for
high
normal highand
and
cancer
predictive
cancer cells
cells and
capability
and 94%
94% offor
flux
for adenomas and
distributions.
adenomas and adenocarcinomas
adenocarcinomas and
and squamous
squamous cell
cell neoplasms,
neoplasms, indicating
indicating high
high
predictive
Copyright
cancer
predictive cellscapability
© 2022
and
capability The
94% of
of
flux
Authors.
for
flux
distributions.
adenomas This is
and
distributions. an open access
adenocarcinomas article under
and the
squamous CC BY-NC-ND
cell neoplasms, license indicating high
Subtyping
Keywords:capability
predictive of and distributions.
staging, lung cancer, modeling, bioinformatics, genome scale metabolic
(https://creativecommons.org
predictive capability
Subtyping of flux
flux
and /licenses/by-nc-nd/4.0/)
distributions.
Keywords:
model,
Keywords: personalized
Subtyping and staging,
medicine, machine
staging, lung cancer,
learning,
lung cancer, modeling,
statistical
modeling, bioinformatics,
analysis, biological network
bioinformatics, genome
genome models scale
scale metabolic
metabolic
Keywords:
model,
Keywords: Subtyping
personalized
Subtyping and
medicine,
and staging,
machine
staging, lung cancer,
learning,
lung cancer, modeling,
statistical
modeling, bioinformatics,
analysis, biological
bioinformatics, genome
network
genome scale
scale metabolic
models metabolic
model, personalized
Keywords: Subtyping medicine,
and machine
staging, learning,
lung cancer, statistical
modeling, analysis, biological network
bioinformatics, genome models
scale metabolic
model,
model, personalized
personalized medicine,
medicine, machine
machine learning,
learning, statistical

statistical analysis,
analysis, biological
biological network
network models
models
model, personalized medicine, machine learning, statistical  tumour analysis,
cells shownbiologicalto use network
lactic modelsacid fermentation instead of
1. INTRODUCTION 
 tumour
oxidative cells shown
phosphorylationto use lactic
due acid fermentation
to fermentation instead
the increasedinstead energy of
1. tumour cells shown to use lactic acid of
1. INTRODUCTION
INTRODUCTION  tumour
oxidative
tumour
demand cells
cells
caused shown
phosphorylation
shown by toto
to use
use lactic
due
lactic
fastuseproliferation acid
to
acid fermentation
the increased
fermentation
rate increased
(Warburg, instead
energy
instead
1956). of
of
Mathematical models 1.
1. key tools to investigate underlying oxidative
INTRODUCTION
are
INTRODUCTION tumour
oxidative cells phosphorylation
shown
phosphorylation
due acid
lactic
due
to fermentation
to
the
the increased
energy
instead
energy of
Mathematical models 1. INTRODUCTION
are key tools to investigate underlying demand
oxidative
GSMMs caused
became by
phosphorylationfast
an proliferation
importantdue to
tool rate
the for (Warburg,
increased
simulation 1956).
energy
and
mechanisms ofmodels
Mathematical complex aremetabolic
key toolsdiseases including
to investigate cancer. demand
underlying oxidative
demand
caused by fast proliferation
phosphorylation
caused by fast due to the
proliferation
rate increased
rate
(Warburg, energy
(Warburg,
1956).
1956).
Mathematical
mechanisms ofmodels
complex are key toolsdiseases
metabolic to investigate
including underlying
cancer. GSMMs
demand
analysis became
causedsuchby
ofbecame an
fast important
proliferation toolrate for simulation
(Warburg, and
1956).
Mathematical
Previous
mechanisms
Mathematical of models
efforts complex
models are key
key tools
on construction
aremetabolic to
of
toolsdiseases investigate
to human genome
including
investigate scale GSMMs
underlying
cancer.
underlying demand
GSMMs caused
became bymetabolic
an
fastimportant
an
alterations.
proliferation
important
tool
tool rateUsing
for
for
this method,
simulation
(Warburg,
simulation
and
1956).
and
mechanisms
Previous of
efforts complex
on metabolic
construction diseases
of human including
genome cancer.
scale analysis
GSMMs of such
became
tissue/disease/condition metabolic
an important
specificalterations.tool
models Using
for
can this
simulation
be method,
obtained and
mechanisms
models
Previous
mechanismsand of
efforts
of complex
integration metabolic
of “-omic”
on construction
complex metabolic diseases
data including
of human
diseases allows
genome
including scale analysis
cancer.
building
cancer. GSMMs
analysis
ofbecame
of
such metabolic
such
tissue/disease/condition
an important
metabolic specific
alterations.
alterations.tool for
models
Using
Using
can be
this method,
simulation
this and
method,
obtained and
Previous
models
Previous andefforts
efforts
patient/context/tissue on
integration
on construction
of “-omic”
construction
specific “-omic” of
of human
data allows
human
constraint-based genome
allows
genome
models scale
building
scale
which analysis
flux of such
distributions
tissue/disease/condition metabolic
at different
specificalterations.
conditions
models Using
can can
be this method,
be analysed
obtained and
models
Previous and integration
efforts on construction of data
of human genome building analysis of
scale tissue/disease/condition such metabolic alterations. Using this method,
models
very and
usefulintegration
patient/context/tissue
are in predicting of
specific “-omic” data
constraint-based allows
models building
which
state. flux
flux
(Reed, distributions
tissue/disease/condition
2012; Schmidt
distributions at etspecific
at different
different
specific
al., 2013), models
conditions
models even,
conditions
can
can be
can
be obtained
canpersonal
can be analysed
be fluxomeand
analysed
be obtained and
models
models and
and integration
patient/context/tissue
integration
patient/context/tissue of flux
of
specific
specific
distributions
“-omic”
“-omic” data
data allows
constraint-based
constraint-based
atmodels
allows steady
models
building
which
building
which
tissue/disease/condition
flux
(Reed,
flux distributions
2012;
distributions Schmidt at
at et
specific
different
al.,
different 2013),
models
conditions
even,
conditions can obtained
be
personal
can be
and
analysed
fluxome
analysed
are
Flux
are very useful
distributions in
patient/context/tissue predicting
at
very useful in predicting specific
a certain flux distributions
constraint-based
condition
flux can at
be
distributions atmodels steady
models
obtained state.
which
steadywhich by
state. flux profile
(Reed, of
2012;the
distributions patient
Schmidt at canet be determined
al.,
different 2013),conditions and
even, reporter
personal
can be reactions
fluxome
analysed
patient/context/tissue specific constraint-based (Reed, 2012; Schmidt 2013),
are
Flux
are very useful
distributions
very useful in
in predicting
at aa (FBA),
certain flux distributions
condition can be at
be steady
obtained by profile
state. profile of
(Reed,
specific 2012;
of the
to patient
Schmidt
the cancer
patient canet be
subtype
canet be
al.,
al., can beeven,
determined
2013),
determined and
even, personal
reporter
personal
identified
and
fluxome
reactions
fluxome
reporter(Ghaffari,
reactions
Flux
are
Flux
Balance
distributions
very
distributions in predicting
usefulAnalysis at
at a
certain
predicting
certain
flux distributions
which
condition
flux
condition
optimizes
distributionscan
can
at
be
steady
atflux rates
obtained
steady
obtained
state.
for
by
state.
by
(Reed,
profile
specific
profile
2012;
of
of the
to
the
Schmidt
patient
cancer
patient canetbe
subtype
can be
2013),
al.,determined
can
determined be
even,
and
and
personal
reporter
identified
reporter
fluxome
reactions
(Ghaffari,
reactions
Flux
given Balance
objective
Flux Balance Analysis
distributions at a
(Ghaffari,
Analysis (FBA),
certain
(FBA), which
condition
Mardinoglu
which optimizes
can
and
optimizes be flux rates
obtained
Nielsen,
flux for
rates forby
2015). Mardinoglu
specific
profile of to
the and
cancer
patient Nielsen,
subtype
can be 2015).
can
determined be Transcriptome
identified
and reporter data,
(Ghaffari,
reactions
Flux distributions
Balance Analysisat a certain
(FBA), condition
which can
optimizes be obtained
flux rates by
for specific
Mardinoglu
specific to
to cancer
and
cancer subtype
Nielsen,
subtype can
2015).
can be
be identified
Transcriptome
identified (Ghaffari,
data,
(Ghaffari,
given
Flux objective
Balance
Typically,
given the
objective (Ghaffari,
Analysis
solution (FBA),
space
(Ghaffari, Mardinoglu
of which
FBA
Mardinoglu is and
optimizes
not Nielsen,
unique
and Nielsen,flux 2015).
rates
because for
of
2015). assuming
Mardinoglu
specific tothat gene
and
cancer expression
Nielsen,
subtype canlevels
2015). be are related
Transcriptome
identified to enzyme
data,
(Ghaffari,
Flux
given Balance
objective Analysis
(Ghaffari,(FBA), which optimizes
Mardinoglu and flux rates
Nielsen, for Mardinoglu
2015). assuming
Mardinoglu that and
gene
and Nielsen,
expression
Nielsen, 2015).
levels
2015). Transcriptome
are related
Transcriptome to data,
enzyme
data,
Typically,
given
the the
objective
high
Typically, solution
degrees
the space
(Ghaffari,
solutionof freedom.
space of FBA
Mardinoglu
of FBATo is
is not
and
address
not unique
Nielsen,
this
unique because
because of
2015).
challenge, of levels,
assuming
Mardinoglu can be
that andused
gene to
Nielsen,constrain
expression 2015). GSMMs.
levels are In
relatedthis
Transcriptome to context,
enzyme
data,
given objective (Ghaffari, Mardinoglu andunique
Nielsen, 2015). assuming
levels, can that
be gene
used expression
to constrain levels are related to enzyme
Typically,
the
Typically,
different
the high
the
high solving
the solution
degrees
solution
degrees methods
of
space
of freedom.
freedom.
space of
of
such FBA
FBA
asToFlux
To
is
is not
address
not this because
unique
Variability
address this challenge,
because
Analysis
challenge,
of
of assuming
most
assuming canthat
levels,commonly be gene
that used
used
gene expression
to methods
constrain
expression toGSMMs.
levels
GSMMs.
levels are In
In this
are related
obtain flux
related to
this context,
enzyme
distributions
context,
to context,
enzyme
Typically,
the high the solution
degrees of space
freedom. of FBATo is not
address unique
this because
challenge, of levels,
most
levels, can
commonly
can be
be used
used
used to
to constrain
methods
constrain to GSMMs.
obtain
GSMMs. In
flux
In this
distributions
this context,
different
the
(FVA)high solving
which
different degrees
solving methods
allows of analysis
methods such
freedom.
such of as
asTo
all Flux Variability
address
solutions
Flux this
in
Variability the Analysis
challenge,
solution
Analysis correlated
most
levels, commonly
can to transcriptome
be used
used to data
methods
constrain were
to Lee12
obtain
GSMMs. (Chung
flux
In this and Lee,
distributions
context,
the high solving
different degreesmethodsof freedom.such asToFlux address this challenge,
Variability Analysis most
most commonly
correlated
commonlyto used
transcriptome
used methods
data
methods to
were
to obtain
Lee12
obtain flux
(Chung
flux distributions
and Lee,
distributions
(FVA) which
different
(FVA) solving
space (Gudmundsson
which allows analysis
methods
allows analysissuch
and such of as
Thiele, all solutions
Flux
2010)
ofasallFlux in
Variability
solutions the solution
Analysis
andinparsimonioussolution most
the Analysis 2012)
correlated or
commonlyto E-flux
transcriptome (Colijn
used methods data et al.,
were
to obtain 2009)
Lee12 flux Additionally,
(Chung and Lee,
distributions
different
(FVA) solving
which methods
allows analysis of all Variability
solutions in the solution correlated
2012)
correlated or to
to transcriptome
E-flux
transcriptome (Colijn data
data et were
al.,
were Lee12
2009)
Lee12 (Chung
(Chung and
and Lee,
Additionally,
Lee,
space
(FVA) (Gudmundsson
which
FBA (pFBA) allows and
analysis Thiele,
of all 2010)
solutionsand in parsimonious
the solution context/tissue/patient
2012) or E-flux specific
(Colijn models can
2009) be generated
Additionally, by
space
(FVA)
space which which
(Gudmundsson
(Gudmundssonallowscalculates
and Thiele,
analysis
and
most
of
Thiele,
efficient
all 2010)
solutions
2010)
and
and
flux distribution
solution correlated
inparsimonious
the
parsimonious 2012)
2012) or toE-flux
transcriptome
context/tissue/patient
or E-flux (Colijn
specific
(Colijn
et
dataetwere al.,
al.,
models
Lee12
2009)
can
2009)
(Chung
be
and Lee,
Additionally,
generated
Additionally, by
FBA
space (pFBA)
through
FBA which
(Gudmundsson
the
(pFBA) objective calculates
and
(Lewis
which calculates most
Thiele,
et efficient
2010)
al.,
most 2010)
efficient and
2010). flux distribution
parsimonious
Integration of
distribution 2012)
fluxparsimonious removing unused
context/tissue/patient
or E-flux (Colijnpathways
specific or et reactions
etmodelsal.,
models
al., cancanusing
2009) be methods
generated such
Additionally, by
space
FBA (Gudmundsson
(pFBA) which and Thiele,
calculates most efficient and flux distribution context/tissue/patient
removing unused
context/tissue/patient specific
pathways
specific or reactions
models can be
using
be generated
methods
generated by
such
by
through
FBA
omic the
(pFBA)
throughdata toobjective
which
genome
the objective (Lewis
calculates
scale
(Lewis most et
most al.,
metabolic
et al., 2010).
efficient Integration
flux
models
2010).flux distribution
is
Integrationanotherof
of removingas GIMME
removing (Becker
unused
context/tissue/patient and
pathways Palsson,
specific or 2008),
reactions
models canINIT
using (Agren
methods et
be generatedsuch al.,
such
by
FBA (pFBA) which calculates efficient distribution unused pathways or reactions using
through
omic
through
method
omic data
data
the
the to
to to objective
objective
decrease
genome
(Lewis
genomesolution
(Lewis
scale
et
scale metabolic al.,
metabolic
space.
et al., 2010).
2010).
Due toIntegration
models
models is ease
Integration
theis another
another
of
of as
as GIMME
removing
2012)
GIMME
removing or iMAT (Becker
unused
unused (Zur,
(Becker and
pathways
Ruppin
and
pathways
Palsson,
or
orand
Palsson, 2008),
reactions
2008), using
Shlomi,
reactions 2010).methods
INIT
INIT
using
(Agren et
methods
(Agren
methods
etsuch
al.,
al.,
such
through
omic datathe toobjective
genome (Lewis
scale et al.,
metabolic 2010).
models Integration of as GIMME
2012) or iMAT
iMAT (Becker
(Zur, andand
RuppinPalsson,
and Shlomi,2008),
Shlomi, INIT (Agren et
2010).(Agren et al., al.,
method
omic data
collecting tocancer
decrease
to genome
relatedsolution
scale space. Due
metabolic
high throughput Due models
omic theis
to data, is another
ease
anotherof 2012) as GIMME or (Becker
(Zur, Ruppin Palsson,
and 2008),
2008), INIT 2010).
method
omic
method datato
to
decrease
to genome
decrease
solution
scale
solution
space.
metabolic
space. Due
to
models
to
the
theis fluxome
ease
another
ease
of
of
as GIMME
2012)
Machine
2012) or
or iMAT
iMAT
(Becker
learning(Zur,
(Zur,
and
Ruppin
became
Ruppin
Palsson,
and
and a Shlomi,
popular
Shlomi,
INIT
tool(Agren
2010).
2010).
et al.,
for cancer
collecting
method
analysis
collecting tocancer
decrease
methods
cancer related
related high
solution
based on
high throughput
space.
integration
throughput Dueomic
of
omic to data,
the
omic
data, fluxome
ease
data
fluxome of
to 2012) or iMAT (Zur, Ruppin and Shlomi, 2010).
method tocancer
decrease solutionthroughput
space. Due to data,
the fluxome
ease of Machine Machine learning learning became aa popular
popular tool forfordatacancer
cancer
collecting
analysis
collecting
genome
analysis methods
cancer
scale
methods
related
basedhigh
related
metabolic
based on
high integration
throughput
models
on (GSMMs)
integration
omic
of omic
omic
of omic
data,
gained data
fluxome
datamoreto researchers
to Machine
for accurate
learning
became
became
classification
aafrompopular
of complex
tool
tool for
sets.
cancer
collecting
analysis cancer
methods related high
based complex
on throughput
integration omic
of omic data,
omic fluxome
data researchers
Machine
They
to researchers simply for
learning accurate
“learn”
for accurate became
a classification
model
classification popular of
providedcomplex
tool
data
of complex fordata
to sets.
cancer
predict
genome
analysis
importance
genome scale
methods
in
scale metabolic
based
studying
metabolic models
on
models (GSMMs)
integration
diseases.
(GSMMs) of gaineddata
Tumour
gained moreto
cells
more Machine
researchers learning
for accuratebecame a popular
classification of tool fordata
complex data
sets.
cancer
sets.
analysis methods based on integration of omic data to the They simply
researchers
future “learn”
for accurate
observations. a model from
classification
Support provided
Vector of data
complex to predict
data sets.
genome
shown toscale
importance
genome
importance scale
have
in
metabolic
in altered
metabolic
studying
models
studyingmetabolism
complex
models
complex
(GSMMs)
diseases.
(GSMMs)
and the most
diseases.
gained
Tumour
gained
Tumourcommonlymore
cells They
more
cells They
simply
researchers
simply
“learn”
for accurate
“learn”
a model
a model
from
classification
from ofMachines
provided
provided
data to
complex
data
(SVM)
predict
data
to
is
sets.
predict
genome
importance scale
in metabolic
studying models
complex (GSMMs)
diseases. gained
Tumour more
cells the
one
They
the future
of these
simply
future observations.
commonly
“learn”
observations. a Support
used
model
Support fromVector
classification Machines
provided
Vector methods.
data
Machines (SVM)
to It is
predict
(SVM) isa
is
shown
defined
shown toexample
importance
to have
in
have altered
studying metabolism
complex
is the metabolism
altered Warburg and
effect.
and thethis
diseases.
In
the most
mostTumourcommonly
phenomenon,
commonly cells They
the simply
future “learn” a model
observations. Support from provided
Vector data to
Machines predict
(SVM) isa
importance in studying complex diseases. Tumour cells one
the of
future these commonly
observations.
commonly Support used classification
Vector methods. It is
shown
defined
shown
defined
to
to have
example
have
example
altered
the metabolism
is the
altered
is Warburg effect.
metabolism
Warburg
and
effect.
and the
In
the
In
most
this
most
this
commonly
phenomenon,
commonly
phenomenon,
one
the of these
one future
of these
these observations.
commonly Support Vector Machines
used classification
used classification
classification
methods.
Machines
methods.
(SVM)
(SVM)It is isa
is isaa
It is
shown
defined to have
example altered
is the metabolism
Warburg and
effect. the
In most
this commonly
phenomenon, one of commonly used methods. It
defined one of these commonly used classification methods. It is a
defined example
2405-8963 example is the Warburg
Copyrightis©the 2022Warburg
The Authors.
effect.
effect. In this phenomenon,
ThisInis this phenomenon,
an open access article under the CC BY-NC-ND license.
Peer review under responsibility of International Federation of Automatic Control.
10.1016/j.ifacol.2022.09.168
642 Ezgi Tanıl et al. / IFAC PapersOnLine 55-20 (2022) 641–646

supervised machine learning algorithm and designed to equipped with COBRA Toolbox (Heirendt et al., 2019) and
determine a decision boundary (optimal hyper plane) between Gurobi 9.5 for flux balance analysis.
two or more data categories that separates the data points
linearly which maximizes the margin to decrease the
classification error (Huang et al., 2018). Numerous
applications of SVM in cancer classification were present in
the literature for different types of cancer using such as
transcriptome (Golub et al., 1999; Furey et al., 2000; Moler,
Chow and Mian, 2000; Li, Zhang and Ogihara, 2004), DNA
methylation (Model et al., 2001; Kim, 2016; Yang et al.,
2017), proteome (Tyanova et al., 2016), copy number
variations (Rapaport, Barillot and Vert, 2008), single
nucleotide polymorphisms (Vural, Wang and Guda, 2016)
and even multi-omic data sets (Kim et al., 2017; Lin and
Lane, 2017). Recently, Xie et al. published a study which
analyses different ML algorithms including SVM from for
early diagnosis of lung cancer using metabolome data (Xie et
al., 2021). Fig. 1. Workflow of the study.
Lung cancer is the main cause of cancer related deaths, with
more than 1.3 million deaths per year, especially for men
2.1 Data Acquisition and Differential Expression Analysis
(17% of the total new cancer diagnosis and 23% of the cancer
related deaths) (Jemal et al., 2011). 5 year survival rate of
Normal (108 samples) and lung cancer high-throughput
lung cancer is around 15% due to the difficulties in
sequencing (htseq-count) data, for both adenoma and
differentiation of lung cancer from other pulmonary diseases.
adenocarcinoma (535 samples) and squamous cell neoplasms
Survival rate may increase to 85% if early diagnosis is
(502 samples) were collected from The Cancer Genome Atlas
possible (Callejon-Leblic et al., 2016). Therefore, diagnosis
Program (TCGA) Research Network
of lung cancer is crucial for application of effective,
(https://www.cancer.gov/tcga). Raw counts were filtered and
personalized treatment. Additionally, accurately determining
normalized, differential expression analysis was performed to
subtype and stage of the cancer using less invasive methods
identify genes that are expressed in significantly different
may allow development of effective and patient specific
levels between normal and tumor tissues (N-C) and adenoma
treatment strategies, even allowing screening programs nation
and adenocarcinomas and squamous cell neoplasms (A-S)
or population wide. Numerous subtyping and staging reports
using DESeq2 (Love, Huber and Anders, 2014). Genes with
were available in literature which are based on differential
absolute log-2 ratio of gene expression values between 2
gene expression, mutations, copy number variation and
groups higher than 2 (|log2FoldChange| > 2) with p-value
metabolite excretion (Cai et al., 2015; Chan and Hughes,
lower than 0.01 were considered as significantly different
2015; Wikoff et al., 2015; Callejon-Leblic et al., 2016;
between two conditions. Gene Ontology (GO) analysis (Yu et
Podolsky et al., 2016; Moreno et al., 2018; Xie et al., 2021).
al., 2012) was performed to identify the biological processes
Although genetic biomarkers such as PIK3CA, TP53, PTEN,
involving significantly expressed genes.
KRAS and EGFR were associated with lung cancer, the
expression profiles of the genes are different for each subtype
and stage. 2.2 Integration of Transcriptome Data to GSMM and Flux
Balance Analysis (FBA)
In this study, early stage lung cancer transcriptome data for
normal and 2 lung cancer subtypes (adenoma and Normalized transcriptome data using DESeq2, were
adenocarcinomas (AD) and squamous cell neoplasms (SC)) integrated to Human Recon3D (Brunk et al., 2018) GSMM
from literature are integrated to human genome scale version published by Lewis et al. (Lewis et al., 2021)
metabolic model (Recon3D). Differential flux and expression containing 13634 reactions and 8457 metabolites (S in (1)
analyses were performed to identify reporter genes and fluxes (13634x8457)) which has corrected energy generating cycles,
for each subtype. SVM algorithm was used to classify using E-flux (Colijn et al., 2009) method, based on the
normal, AD and SC samples with minimum number of assumption that fluxes are correlated to the expression of
predictors and highest accuracy. Finally, functional analysis corresponding genes. According to the transcriptome data,
such as Gene Ontology analysis, pathway enrichment for reaction expression values (aj, bj in (1)) that will be used as
hunting hallmarks of cancer were performed. flux bounds that were mapped benefiting Gene-Protein-
Reaction (GPR) association rules. GPR rules are defined in
2. METHOD the model describing reactions catalysed by enzymes encoded
by a single gene and multiple genes encoding enzymatic
The workflow of the study is presented in Fig 1. Differential
subunits and isozymes. Overall expression levels for fluxes
expression and Gene Ontology (GO) analysis were performed
were calculated using ‘minSum’ option which uses the
in R. Calculations were performed in MATLAB r2018a
minimum expression value for enzymes containing subunits
and sum of expression values of all isozymes and then set as
 Ezgi Tanıl et al. / IFAC PapersOnLine 55-20 (2022) 641–646 643

upper and lower bounds. FBA was performed to obtain flux chromosome assembly etc.) or signalling which is parallel to
distributions (v in (1)) by maximizing growth, which can be the hallmarks of cancer (Hanahan and Weinberg, 2011).
used as objective for cancer cells because their main aim is
proliferate and invade other tissues (Damiani et al., 2017;
Nilsson and Nielsen, 2017).
a)

(1)

2.3 Differential Flux Analysis and Pathway Enrichment

Differential flux analysis was performed using non-

parametric Wilcoxon rank sum test and FDR correction, by
comparing distributions and expression of each flux for
normal and tumour samples (N-C) and adenomas and
adenocarcinomas and squamous cell neoplasms (A-S) (Nanda
and Ghosh, 2020). Fluxes that have p-values lower than 0.05
were determined to be significantly different between two
conditions. Pathway enrichment was performed using
hypergeometric test employing subsystems defined in model.
This test is commonly used to determine whether a property
is overly represented or not in selected sample from
population. Subsystems that have p-values lower than 0.1
were determined to be overrepresented.

2.4 Linear SVM Classification for N-C and A-S

A cascaded-tree like subtyping approach was performed by

first classifying normal (N) and tumour cells (C). Then b)
tumour cells were classified into subtypes. Using fluxome
data randomly selected 80 normal samples and 850 cancer
samples (420 samples for adenoma and adenocarcinoma
(AD) and 430 samples for squamous cell neoplasms (SC))
were selected to train SVM model with 10 fold cross-
validation. Remaining data were used to test the model.
Reactions shown to carry significant fluxes between C/N and
SC/AD (p<0.05, |log2FoldChange|>1) were chosen as
potential predictors for classification. Then, to decrease the
number of predictors, effect of each predictor on validation
accuracy were calculated. This process is applied for both N-
C and A-S classification to obtain models that give maximum
accuracy with minimum number of predictors.

3. RESULTS AND DISCUSSION

3.1 Differential Expression and GO Analysis

Differential expression analysis was performed to determine

the genes that have significant difference between normal and
cancer cells and adenomas and adenocarcinomas and
squamous cell neoplasms. Out of 60483 genes, 5998 and
4074 genes were found to have significantly different
expression levels between N-C and AD-SC, respectively. GO
analysis results for the gene sets were shown in Fig 2 which
indicates that, mostly genes that are significantly different in
each case belong to processes of either cellular structure
formation, cell cycle (DNA replication, packaging, Fig. 2. GO analysis results for a) N-C and b) AD-SC.
644 Ezgi Tanıl et al. / IFAC PapersOnLine 55-20 (2022) 641–646
Additionally, since these profiles do not represent the disease
state only, unrelated mechanisms such as sex differentiation,
regulation of water loss via skin were found to be significant.
In conclusion, although transcriptomic analysis gives
potential genetic biomarkers, these are not lung cancer
specific and they cannot be measured easily.
3.2 Differential Flux Analysis and Pathway Enrichment
Wilcoxon rank sum test and hypergeometric test were
performed to analyse statistical significance of the differences
flux distributions and altered subsystems between different
cell types. Pathway enrichment results for normal and cancer
cells and subtypes and the corresponding p-values were
shown in Fig 3. 292 reactions were found to have
significantly different fluxes between normal and cancer
cells, 103 reactions were found to have significantly different
fluxes between adenoma and adenocarcinomas and squamous
cell neoplasms.
a)
b)
Fig. 3. Pathway enrichment results for a) cancer-normal
groups b) subtypes. Dot sizes are proportional to number of
fluxes involved in the subsystem.
Model currently represents tumour microenvironment as a
single cell since transcriptome data is obtained from a tumour
tissue sample, not the tumour cell itself. Cancer progress can
be characterized by shifting towards oxidative stress due to
uncontrolled growth. In this context, elevated flux through
reduced glutathione exchange can represent the protection
mechanisms of cancerous cells from reactive oxygen species.
Changes in glycolysis and citric acid cycle are another
mechanism reported to be affected during reprogramming of
metabolism in cancer cells (Warburg effect). Additionally,
amino acid metabolism were reported to be altered since they
are used as growth factors such as glutamate (Stepulak et al.,
2014) and providing additional energy for uncontrolled
proliferation. Oxidative stress in cancerous cells also affects
lipid metabolism and oxidation pathways which are mainly
involved in signalling and proliferation and metastasis.
Inositol and fatty acid metabolisms were reported to be
altered in highly metastatic cancer cell lines which is in line
with our results.
Another hallmark of cancer is tissue inflammation. Tissue
inflammation in lung cancer can be mediated by leukotrienes
which are biologically active eicosanoid lipid mediators
secreted by immune cells in tumour microenvironment.
Elevated levels leukotriene B4 (LTB4) was reported to be
observed in lung cancer. Also LTB4 enhances angiogenesis
and blood vessel permeability (Salvador et al., 2017; Tian et
al., 2020).
Nucleotide and cofactor metabolisms are other altered
metabolisms in cancer cells. They have a key role in DNA
synthesis and repair and increasing proliferation rate
increases the nucleotide and NAD+ demand. In this context,
inosine, adenosine, nicotinate, riboflavin concentrations were
reported to be increased in various cancer cells (Hsu and
Sabatini, 2008).
Overall, considering 12 hallmarks of cancer and previous
transcriptome and metabolome markers published in
literature (Hsu and Sabatini, 2008), it was shown that model
reflected the biochemistry of the tumour environment
qualitatively.
3.3 Linear SVM classification
Classification capability of significantly different fluxes were
tested with linear SVM classification. In classification of
normal and cancer cells, accuracy of 0.991 was obtained
using all 292 reactions as predictors. When number of
predictors increased one by one choosing the ones that have
highest validation accuracy, final accuracy of 0.995 was
obtained for 7 predictors as shown in Fig 4.a. For adenomas
and adenocarcinomas and squamous cell neoplasms
classification, 0.931 accuracy was obtained when all 103
fluxes were used as predictors. On the other hand, 0.94
accuracy was obtained using 13 predictors as shown in Fig
4.b. Overall, classification results shown that model flux
predictions enables adequate classification. Also, flux
predictions were shown to have potential to be used for
enhancing accuracy and precision of classifications
performed using transcriptome data.
 Ezgi Tanıl et al. / IFAC PapersOnLine 55-20 (2022) 641–646 645
a)
b)
Fig. 4. Classification accuracies calculated cumulatively for
a) cancer-normal groups b) subtypes.
4. CONCLUSIONS
Personalized medicine requires correct subtyping of complex
diseases such as lung cancer using methods less invasive as
possible. This application illustrates genome scale metabolic
model-based subtyping among different cell types or disease
conditions. By incorporating models of metabolism allows
diagnosis with non-invasive methods, by transforming
transcriptome level information to more accessible metabolite
level information.
5. ACKNOWLEDGEMENTS
This work is funded by Health Institutes of Turkey (TÜSEB)
project no: 2019-TA-01-4589.
REFERENCES
Agren, R. et al. (2012) ‘Reconstruction of genome-scale
active metabolic networks for 69 human cell types and
16 cancer types using INIT’, PLoS Comput Biol, 8(5), p.
e1002518.
Becker, S. A. and Palsson, B. O. (2008) ‘Context-specific
metabolic networks are consistent with experiments’,
PLoS Comput Biol, 4(5), p. e1000082.
Brunk, E. et al. (2018) ‘Recon3D enables a three-dimensional
view of gene variation in human metabolism’, Nature
biotechnology, 36(3), p. 272.
Cai, Z. et al. (2015) ‘Classification of lung cancer using
ensemble-based feature selection and machine learning
methods’, Molecular BioSystems, 11(3), pp. 791–800.
Callejon-Leblic, B. et al. (2016) ‘Metabolic profiling of
potential lung cancer biomarkers using bronchoalveolar
lavage fluid and the integrated direct infusion/gas
chromatography mass spectrometry platform’, Journal of
proteomics, 145, pp. 197–206.
Chan, B. A. and Hughes, B. G. M. (2015) ‘Targeted therapy
for non-small cell lung cancer: current standards and the
promise of the future’, Translational lung cancer
research, 4(1), p. 36.
Chung, B. K. S. and Lee, D. Y. (2012) ‘Computational codon
optimization of synthetic gene for protein expression’,
BMC Systems Biology, 6(1), p. 1.
Colijn, C. et al. (2009) ‘Interpreting expression data with
metabolic flux models: predicting Mycobacterium
tuberculosis mycolic acid production’, PLoS Comput
Biol, 5(8), p. e1000489.
Damiani, C. et al. (2017) ‘A metabolic core model elucidates
how enhanced utilization of glucose and glutamine, with
enhanced glutamine-dependent lactate production,
promotes cancer cell growth: The Warburg effect’, PLoS
computational biology, 13(9), p. e1005758.
Furey, T. S. et al. (2000) ‘Support vector machine
classification and validation of cancer tissue samples
using microarray expression data’, Bioinformatics,
16(10), pp. 906–914.
Ghaffari, P., Mardinoglu, A. and Nielsen, J. (2015) ‘Cancer
metabolism: a modeling perspective’, Frontiers in
physiology, 6, p. 382.
Golub, T. R. et al. (1999) ‘Molecular classification of cancer:
class discovery and class prediction by gene expression
monitoring’, science, 286(5439), pp. 531–537.
Gudmundsson, S. and Thiele, I. (2010) ‘Computationally
efficient flux variability analysis’, BMC Bioinformatics,
11(2), pp. 2–4. doi: 10.1186/1471-2105-11-489.
Hanahan, D. and Weinberg, R. A. (2011) ‘Hallmarks of
cancer: the next generation’, cell, 144(5), pp. 646–674.
Heirendt, L. et al. (2019) ‘Creation and analysis of
biochemical constraint-based models using the COBRA
Toolbox v. 3.0’, Nature protocols, 14(3), pp. 639–702.
Hsu, P. P. and Sabatini, D. M. (2008) ‘Cancer cell
metabolism: Warburg and beyond’, Cell, 134(5), pp.
703–707.
Huang, S. et al. (2018) ‘Applications of support vector
machine (SVM) learning in cancer genomics’, Cancer
Genomics-Proteomics, 15(1), pp. 41–51.
Jemal, A. et al. (2011) ‘Global cancer statistics.’, CA: a
cancer journal for clinicians, 61(2), pp. 69–90.
Kim, S. (2016) ‘Weighted K-means support vector machine
646 Ezgi Tanıl et al. / IFAC PapersOnLine 55-20 (2022) 641–646

for cancer prediction’, Springerplus, 5(1), pp. 1–11. models of cellular metabolism developed from
Kim, S. et al. (2017) ‘Meta-analytic support vector machine metabolomics and expression data’, Bioinformatics,
for integrating multiple omics data’, BioData mining, 29(22), pp. 2900–2908.
10(1), pp. 1–14. Stepulak, A. et al. (2014) ‘Glutamate and its receptors in
Kuepfer, L., Sauer, U. and Blank, L. M. (2005) ‘Metabolic cancer’, Journal of neural transmission, 121(8), pp. 933–
functions of duplicate genes in Saccharomyces 944.
cerevisiae’, Genome research, 15(10), pp. 1421–1430. Tian, W. et al. (2020) ‘Leukotrienes in Tumor-Associated
Lewis, J. E. et al. (2021) ‘Personalized genome-scale Inflammation’, Frontiers in Pharmacology, 11, p. 1289.
metabolic models identify targets of redox metabolism in Tyanova, S. et al. (2016) ‘Proteomic maps of breast cancer
radiation-resistant tumors’, Cell Systems, 12(1), pp. 68– subtypes’, Nature communications, 7(1), pp. 1–11.
81. Vural, S., Wang, X. and Guda, C. (2016) ‘Classification of
Lewis, N. E. et al. (2010) ‘Omic data from evolved E. coli breast cancer patients using somatic mutation profiles
are consistent with computed optimal growth from and machine learning approaches’, BMC systems
genome‐scale models’, Molecular systems biology, 6(1), biology, 10(3), pp. 263–276.
p. 390. Warburg, O. (1956) ‘On the origin of cancer cells’, Science,
Li, T., Zhang, C. and Ogihara, M. (2004) ‘A comparative 123(3191), pp. 309–314.
study of feature selection and multiclass classification Wikoff, W. R. et al. (2015) ‘Metabolomic markers of altered
methods for tissue classification based on gene nucleotide metabolism in early stage adenocarcinoma’,
expression’, Bioinformatics, 20(15), pp. 2429–2437. Cancer prevention research, 8(5), pp. 410–418.
Lin, E. and Lane, H.-Y. (2017) ‘Machine learning and Xie, Y. et al. (2021) ‘Early lung cancer diagnostic biomarker
systems genomics approaches for multi-omics data’, discovery by machine learning methods’, Translational
Biomarker research, 5(1), pp. 1–6. oncology, 14(1), p. 100907.
Love, M. I., Huber, W. and Anders, S. (2014) ‘Moderated Yang, Z. et al. (2017) ‘Classification based on feature
estimation of fold change and dispersion for RNA-seq extraction for hepatocellular carcinoma diagnosis using
data with DESeq2’, Genome biology, 15(12), pp. 1–21. high-throughput dna methylation sequencing data’,
Model, F. et al. (2001) ‘Feature selection for DNA Procedia Computer Science, 107, pp. 412–417.
methylation based cancer classification’, Bioinformatics, Yu, G. et al. (2012) ‘clusterProfiler: an R package for
17(suppl_1), pp. S157–S164. comparing biological themes among gene clusters’,
Moler, E. J., Chow, M. L. and Mian, I. S. (2000) ‘Analysis of Omics: a journal of integrative biology, 16(5), pp. 284–
molecular profile data using generative and 287.
discriminative methods’, Physiological genomics, 4(2), Zur, H., Ruppin, E. and Shlomi, T. (2010) ‘iMAT: an
pp. 109–126. integrative metabolic analysis tool’, Bioinformatics,
Moreno, P. et al. (2018) ‘Metabolomic profiling of human 26(24), pp. 3140–3142.
lung tumor tissues–nucleotide metabolism as a candidate
for therapeutic interventions and biomarkers’, Molecular
oncology, 12(10), pp. 1778–1796.
Nanda, P. and Ghosh, A. (2020) ‘Genome Scale-Differential
Flux Analysis reveals deregulation of lung cell
metabolism on SARS Cov2 infection’, bioRxiv.
Nilsson, A. and Nielsen, J. (2017) ‘Genome scale metabolic
modeling of cancer’, Metabolic engineering, 43, pp.
103–112.
Podolsky, M. D. et al. (2016) ‘Evaluation of machine
learning algorithm utilization for lung cancer
classification based on gene expression levels’, Asian
Pacific Journal of Cancer Prevention, 17(2), pp. 835–
838.
Rapaport, F., Barillot, E. and Vert, J.-P. (2008)
‘Classification of arrayCGH data using fused SVM’,
Bioinformatics, 24(13), pp. i375–i382.
Reed, J. L. (2012) ‘Shrinking the metabolic solution space
using experimental datasets’, PLoS Comput Biol, 8(8), p.
e1002662.
Salvador, M. M. et al. (2017) ‘Lipid metabolism and lung
cancer’, Critical Reviews in Oncology/Hematology, 112,
pp. 31–40.
Schmidt, B. J. et al. (2013) ‘GIM3E: condition-specific