Báo cáo khoa học: A systems biological approach suggests that transcriptional feedback regulation by dual-specificity phosphatase 6 shapes extracellular signal-related kinase activity in RAS-transformed fibroblasts pptx

Kielbasa3, Anja Schramme2, Oleg Tchernitsa2, Jana Keil2, Andrea Solf2, Martin Vingron3, Reinhold Scha¨fer2, Hanspeter Herzel1and Christine Sers2 1 Institute for Theoretical Biology, Humb

transcriptional feedback regulation by dual-specificity

phosphatase 6 shapes extracellular signal-related kinase activity in RAS-transformed fibroblasts

Nils Blu¨thgen1,2, Stefan Legewie1, Szymon M Kielbasa3, Anja Schramme2, Oleg Tchernitsa2, Jana Keil2, Andrea Solf2, Martin Vingron3, Reinhold Scha¨fer2, Hanspeter Herzel1and

Christine Sers2

1 Institute for Theoretical Biology, Humboldt University, Berlin, Germany

2 Laboratory of Molecular Tumor Pathology, Charite´, Universita¨tsmedizin Berlin, Germany

3 Computational Biology, Max Planck Institute for Molecular Genetics, Berlin, Germany

The mitogen-activated protein kinase cascade (MAPK)

activating extracellular signal-related kinase ERK1 and

ERK2 controls crucial cell fate decisions such as

differ-entiation, proliferation and malignant transformation

Quantitative differences in signal strength or signal duration result in specific cell fates, e.g either prolifer-ation or differentiprolifer-ation [1] The activity of ERK1,2 is regulated through a balance of stimulation through

Keywords

dual-specificity phosphatase; mathematical

modelling; mitogen activated protein kinase;

transcriptional feed-back

Correspondence

N Blu¨thgen, Institute of Pathology,

Universita¨tsmedizin Charite´, FORSYS junior

group, Chariteplatz 1, D-10117 Berlin

Fax: +49 30 450 536 909

Tel: +49 30 450 536 134

E-mail: nils.bluethgen@charite.de

Database

The mathematical model described here has

been submitted to the Online Cellular

Systems Modelling Database and can be

accessed at

http://jjj.biochem.sun.ac.za/data-base/bluthgen/index.html free of charge

(Received 10 July 2008, revised 8

November 2008, accepted 8 December

2008)

doi:10.1111/j.1742-4658.2008.06846.x

Mitogen-activated protein kinase (MAPK) signaling determines crucial cell fate decisions in most cell types, and mediates cellular transformation in many types of cancer The activity of MAPK is controlled by reversible phosphorylation, and the quantitative characteristics of MAPK activation determine the cellular response Many systems biological studies have analyzed the activation kinetics and the dose–response behavior of the MAPK signaling pathway Here we investigate how the pathway activity is controlled by transcriptional feedback loops Initially, we predict that MAPK signaling regulates phosphatases, by integrating promoter sequence data and ontology-based classification of gene function From this, we deduce that MAPK signaling might be controlled by transcriptional nega-tive feedback regulation via dual-specificity phosphatases (DUSPs), and implement a mathematical model to further test this hypothesis Using time-resolved measurements of pathway activity and gene expression, we employ a model selection approach, and select DUSP6 as a highly likely candidate for shaping the activity of the MAPK pathway during cellular transformation caused by oncogenic RAS Two predictions from the model were confirmed: first, feedback regulation requires that DUSP6 mRNA and protein are unstable; and second, the activation kinetics of MAPK are ultrasensitive Taken together, an integrated systems biological approach reveals that transcriptional negative feedback controls the kinetics and the extent of MAPK activation under both physiological and pathological conditions

Abbreviations

CREB, cAMP response element binding protein; DUSP, dual-specificity phosphatase; ERK, extracellular signal-related kinase; FDR, false discovery rate; IPTG, isopropyl-thio-b- D -galactoside; IR, inducible RAS; MAPK, mitogen-activated protein kinase; MEK, mitogen-activated protein kinase ⁄ extracellular signal-related kinase kinase; PDGF, platelet-derived growth factor; siRNA, small intefering RNA; SRF, serum response factor.

upstream kinases [MAPK⁄ ERK (MEK)1,2] and

inhibi-tory actions, namely dephosphorylation through

spe-cific phosphatases Most experimental and theoretical

approaches have focused on the biochemical

mecha-nisms and on the spatiotemporal ordering mediating

ERK1,2 activation These approaches lead to the

assumption that simply inhibiting MEK1,2 or ERK1,2

using therapeutic small-molecule inhibitors would be

sufficient to suppress pathway activation and thereby

reverse downstream biological responses such as

immune function, mitogenesis or even malignant cell

growth However, many inhibitors directly targeting

MEK or upstream kinases have produced

unpredict-able cellular and clinical responses [2]

The MAPK signaling network has been investigated

by mathematical modeling for more than a decade

[3–5] The input–output relationships of the MAPK

cascade have been intensively studied by mechanistic

modeling Studies in Xenopus oocyctes have suggested

that the cascade-like structure and double

phosphory-lation of MEK and ERK give rise to a nonlinear

sigmoidal response [6] Moreover, the influence of

post-translational feedback loops on the dynamic

behavior of this signaling cascade has been unveiled,

and it was found that there are positive and negative

feedbacks, depending on the cellular context [7]

Posi-tive feedback has been shown to increase the sensitivity

of the stimulus–response relationships It may even

cause bistability, where the state of signaling may

depend on whether the pathway has been stimulated

earlier [8] In contrast, negative MAPK feedback

allows the MAPK cascade to return to lower activity,

even if upstream signaling persists, and therefore to

adapt to prolonged extracellular stimulation [9] If the

signaling pathway is very sensitive, negative feedback

can bring about oscillations This has been postulated

by Kholodenko for MAPK signaling [10], and was

recently observed experimentally [11]

So far, the consequences of transcriptional feedbacks

in MAPK signaling have not been addressed in detail

Currently, the majority of biological information on

negative regulation of MEK⁄ ERK signaling is derived

from studies on mouse, chicken and zebrafish

develop-ment [12–15]; the relevance in adult animals is less

clear These studies revealed an essential role for the

ERK-specific dual-specificity phosphatase DUSP6 in

development, and showed that it acts downstream of

the fibroblast growth factor receptor to inhibit the

ERK response Previous mathematical models were

focused on the control of ERK activation by

hormones at short time scales of < 60 min, and the

concentrations of the proteins were assumed to be

con-stant and independent of transcriptional changes

However, many important cell fate decisions and cellu-lar transformations are slow processes that require long-term MAPK activation and subsequent altera-tions in gene expression [16,17] Downstream of ERK, numerous transcription factors become activated in sequential transcriptional cascades It is believed that distinct combinations of transcription factors give rise

to a specific cellular response [18,19] In attempts to predict the transcription factors that are functionally involved in certain ERK-dependent processes, even sophisticated methods, including combinatorial approaches or the analysis of phylogenetic conserva-tion of potential regulatory sites, have proven unsatis-factory [20] Therefore, the transcriptional response was only rarely taken into account in modeling approaches addressing ERK1,2 signaling, and the role

of individual transcription factors targeted through MEK⁄ ERK signaling was not included

Here we aim at identifying feedback adaptation mechanisms within the MEK⁄ ERK signaling cascade

by first scanning MAPK target genes for potential functions in MAPK signaling Candidate transcrip-tional feedback loops are then further analyzed using a semiquantitative mathematical model of MAPK signal-ing that incorporates changes in the transcriptome This approach allows us to identify transcriptional feedback loops that may be important in cellular trans-formation and for cell fate decisions

The mathematical model described here has been submitted to the Online Cellular Systems Modelling Database and can be accessed at http://jjj.biochem sun.ac.za/database/bluthgen/index.html free of charge

Results Transcription factors downstream of ERK are predicted regulators of phosphatase function The genome-wide prediction of target genes of a partic-ular transcription factor is far from being reliable [21] Therefore, we developed and validated a function-ori-ented approach to predict target genes responding to ERK activation [22] Instead of screening promoter sequences for transcription factor binding sites and pre-dicting target genes directly, we asked which cellular functions are regulated by a specific transcription factor

or combinations thereof Once a set of defined tran-scription factors was identified, trantran-scriptional targets were predicted and further tested for enrichment in annotated functions as described by gene ontology [22] We applied this algorithm to serum response fac-tor (SRF) and cAMP response element binding protein (CREB), two central transcription factors downstream

of ERK [23] Our algorithm identified the terms

‘protein amino acid dephosphorylation’ and

‘dephos-phorylation’ as the only terms that are significantly

enriched within the group of putative SRF⁄ CREB

target genes (PFDR< 0.05, where FDR is false

discov-ery rate) Therefore, we speculated that phosphatases

might feed back into MAPK signaling Good

candi-dates for such feedback mechanisms are the classical

DUSPs, a family of phosphatases that specifically

dephosphorylate MAPKs [24] Therefore, we collected

evidence that DUSPs are regulated by these two

tran-scription factors A recent

chromatin-immunoprecipia-tion on chip (ChIP-on-chip) experiment demonstrated

that the promoters of DUSP1, DUSP3, DUSP4,

DUSP6 and DUSP11 are directly bound by CREB

[25], suggesting a direct involvement of CREB in their

transcriptional regulation To further confirm

SRF-dependent DUSP regulation experimentally, we tested

the effect of SRF silencing on DUSP4 and DUSP6

expression After transient transfection of

HRAS-trans-formed immortalized human embryonal kidney cells

[26] with two independent small intefering RNAs

(siR-NAs) specifically targeting the SRF gene, we analyzed

SRF, DUSP4and DUSP6 mRNAs by real-time PCR

Transfection of the cells with SRF-specific siRNAs

sup-pressed SRF expression itself, but also that of the two

phosphatase genes, after 96 h (Fig 1) These results

largely confirm our prediction of a strong impact

of SRF on DUSP regulation, and suggest that

ERK might regulate its own activity by inducing

phos-phatases, at least under certain biological conditions

Model selection suggests that DUSP6 is induced

and modulates ERK activity

Having identified putative direct links between

MAP-Ks, the transcription factors SRF and CREB and the

regulation of DUSPs, we aimed at investigating

whether the regulation of these phosphatases

consti-tutes feedback loops that modulate MAPK activation

in vivo Experimental evidence indicated that many

receptor-mediated stimuli cause rapid adaptation and

desensitization of the receptors and of signaling

mole-cules by post-translational modifications and receptor

internalization [27] Thus, it is impossible to distinguish

feedbacks due to the transcriptional regulation of

phosphatases from feedbacks due to receptor

deactiva-tion when the cells are stimulated at the level of the

receptors Consequently, we decided to stimulate the

canonical MAPK cascade using RAS constructs

encoding mutationally activated RAS proteins that

signal constitutively without requiring receptor

activa-tion In immortalized rat fibroblasts, expression of

oncogenic RAS (H-RASV12) elicits prolonged activa-tion of ERK and cellular transformaactiva-tion [17]

To investigate the dynamic implications of a putative DUSP-mediated feedback, we used an inducible onco-genic H-RASV12gene construct controlled by an isopro-pyl-thio-b-d-galactoside (IPTG)-sensitive promoter [28] After addition of IPTG to the medium, the cells express oncogenic RAS [29] We monitored RAS expression and ERK phosphorylation by western blot, and the transcriptional levels of several DUSPs by interrogating custom microarrays [30] and by northern blots in a time-resolved manner (Fig 2A,B) The RAS protein is strongly induced and accumulates through-out the measurement period, whereas ERK is initially strongly activated, then declines, and is subsequently maintained at an intermediate level of activation Among several DUSPs, the relatively unspecific DUSP1 and the ERK-specific DUSP6 show rapid induction Therefore, both are likely candidates for negatively regulating ERK activity and causing the biphasic response of ERK activation To further

DUSP4 mRNA

0 0.4 0.8 1.2

0 0.4 0.8 1.2

0 0.04 0.08 0.12

DUSP6 mRNA

SRF mRNA

SRF 1 SRF 2 SCR 1 SCR 2 SCR 3

siRNA

Fig 1 Expression of DUSP4 and DUSP6 depends on SRF Real-time PCR analysis of SRF, DUSP4 and DUSP6 mRNA expression

96 h following transfection with siRNAs suppressing SRF Two independent SRF siRNAs (SRF_1 and SRF_2) were used, and three scrambled siRNAs (SRC1 ⁄ 2 ⁄ 3) were used as controls Reactions for SRF, DUSP4 and DUSP6 were performed, and the cycle thresh-old (CT) values are depicted All reactions were normalized to relative levels of tubulin as an internal standard.

explore this hypothesis, we quantified time-resolved

lev-els of RAS and phosphorylated ERK proteins by

wes-tern blot, and of DUSP6 and DUSP1 mRNA by

northern blot, and fitted different mathematical models

to the data The induction kinetics of RAS in the

experimental system varied from time-series experiment

to time-series experiment It is therefore important that

we used both mRNA and protein samples from the

same experimental time-series experiments for model

construction and fitting, and did not use the microarray

data, which came from a different experimental run

We applied a model selection process based on the

likelihood ratio test [31] Briefly, for each of the models

investigated, the best fit of the model to the data was

obtained by a maximum likelihood method The

good-ness of fit was quantified by calculating the v2-value,

i.e the sum of the squared differences between data

and model fit divided by the variance of the data A

more complex model can fit better, because it describes

the system better, or because it fits the experimental

error (also called overfitting) In order to discriminate between these two scenarios, we calculated P-values that quantified the probability that a model fits the data better just because the alternative model fits the noise better These P-values were estimated using a Monte Carlo method (for details, see Appendix S2 and [32]) Using this approach, we can determine whether adding additional molecular processes to the model or assuming different mechanisms in the model improves the description of data just because the model has more degrees of freedom, which would lead to model rejec-tion If the new molecular steps are essential to the model, the model will be accepted

We first constructed two mathematical models: one model describing ERK activation with DUSP6-medi-ated ERK dephosphorylation and another model with-out (Fig 3A,B) We found that DUSP6-mediated ERK dephosphorylation is indeed required for the model to properly describe the data, as otherwise the biphasic response in ERK phosphorylation cannot be

ERKPP

ERK

RAS

DUSP6 DUSP6

0 1 2

0 1 2

0 5 10 15 20

Time (h)

0 1 2

IPTG

0 1 2

0 1 2

0 5 10 15 20

Time (h)

0 1 2

western blot

western blot

northern blot

0

Time (h)

DUSP6

DUSP1 DUSP5

DUSP9

Array Northern blot

Fig 2 Model construction from time-series data (A) After induction of oncogenic RAS, several DUSPs are transcriptionally regulated, as detected by microarrays (gray) and northern blots (black) DUSP6, a very specific phosphatase for ERK, is rapidly upregulated after induction (B) Quantified western blot and northern blot time series show that RAS expression increases monotonically over the first 24 h after induc-tion ERK phosphorylation is first increased, and then briefly decreased, followed by a plateau Actin measurements are used for normaliza-tion of RAS signals, and phospho-ERK levels are normalized by total ERK intensities DUSP6 mRNA levels rapidly rise after RAS is induced (C) Schematic representation of the selected mathematical model RAS is induced by IPTG and degraded ERK is phosphorylated as a conse-quence of RAS activation, and phospho-ERK in turn induces DUSP6 mRNA expression DUSP6 is translated into DUSP6 protein, which dephosphorylates ERK In the final model, ERK (de)phosphorylation is assumed to be in quasi-steady state, with nonlinear dependence on RAS and DUSP6 (D) Time-series of the best fit of the final model together with the quantified time-series data from western blots (for ERKpp and RAS), and northern blot (for DUSP6 mRNA).

ERKP ERK RAS

dusp6

ERKPP ERK RAS

dusp6

DUSP6

Model A

without feedback

Model B

with DUSP6

Model D

reduced model

Model E

with ultrasensitivity

ERKPP RAS

dusp6

DUSP6

Ultrasensitive

ERKP RAS

dusp6 DUSP6

Linear

Model reduction explains data similarly well

(P > 0.6)

Model with DUSP6 feedback

explains data better (P < 0.01)

Model fits better with ultrasensitivity (P < 0.05)

ERKP ERK RAS

dusp1 DUSP1

Model C

with DUSP1

DUSP6 feedback explains data

better than with DUSP1 (P < 0.01)

0 1 2

0 1 2

Time (h)

0

1

0

1

2

0

1

2

Time (h)

0

1

2

0

1

2

0

1

2

Time (h)

0

1

2

0 1 2

0 1 2

Time (h)

0 1 2

0 1 2

0 1 2

Time (h)

0 1 2

Fig 3 Model selection procedure The structure and the best fit to the first data points of the five models are shown, as well as the P-val-ues from the likelihood ratio test (A) First, a model without feedback was constructed and fitted to time-series data of RAS protein expres-sion, ERK phosphorylation, and dusp6 mRNA expression This model could not reproduce the biphasic response (B) A model that includes dephosphorylation of ERK explains the data significantly better (C) Fitting the same model to time-series of dusp1 mRNA results in a signifi-cantly worse fit (D) Model B was reduced by quasi-steady-state approximation of ERK activity This reduced model fits the data similarly well (E) Erk activation and deactivation was assumed to be ultrasensitive, and this model fits the data significantly better than model D.

reproduced We also investigated whether DUSP1 can

similarly account for the observed dynamics in ERK

phosphorylation by fitting the model with feedback to

the time course of ERK, RAS and DUSP1 mRNA

(Fig 3C) This model fitted significantly less well,

which suggests that DUSP6 is the important regulator

in the first hours of ERK signaling

Therefore, we chose the model structure shown in

Fig 3B, and investigated whether we can reliably

determine the parameters in the mathematical model

from the experimental data We used a Monte Carlo

approach to define confidence intervals for and

corre-lation coefficients between the parameters The

param-eters describing ERK activation and deactivation

showed large confidence intervals, and were highly

cor-related, which indicates that they are not identifiable

(for details see Appendix S1) This is not too

surpris-ing, as the typical time scale for activation and

deacti-vation of ERK is much smaller than the intervals

between the time points of measurements of ERK

phosphorylation Thus, the detailed activation and

deactivation rates cannot be inferred separately from

our data Moreover, the parameters describing the

impact of ERK phosphorylation on DUSP6 expression

and vice versa were especially highly correlated

There-fore, we reduced model complexity by applying a

quasi-steady-state approximation for phosphorylation

and dephosphorylation of ERK Model selection

shows that the resulting reduced model fits the data

similarly well as the more detailed model (Fig 3D and

Appendix S2)

The model also allows us to investigate whether

ERK activation is responding to upstream events in a

linear or nonlinear manner Mechanistic modeling has

suggested that ERK and MEK respond in an

ultrasen-sitive fashion [6,33,34], but so far this has only been

confirmed for signaling processes in Xenopus oocytes

We modified the model such that ERK activation is

nonlinear with an exponent of 2 This modified model

fitted the data significantly better and allowed us to

describe the biphasic response of ERK more precisely

(Fig 3E) The structure and time-series of the best fit

of this final model is shown in Fig 2C,D

In conclusion, model selection of the time-series data

resulted in two testable predictions First, the two

parameters describing DUSP6 mRNA and protein

decay in the model have a direct biophysical meaning

Both DUSP6 mRNA and protein are estimated to be

rapidly decaying, which can be compared to the

bio-chemical data Second, the model selection predicts

that the activation of ERK is nonlinear As described

in the following, we collected quantitative experimental

measurements to test these model predictions

Model prediction 1 – DUSP6 is unstable at the mRNA and protein levels

Most of the model parameters are given in relative units; thus, they cannot be compared to biochemical measurements However, two parameters in the model have a direct biophysical meaning: the decay rates of DUSP6 protein and mRNA are estimated to be relatively fast, at 3.5 and 0.9 h)1, respectively These values correspond to half-lives of 11 and 46 min for

0 20 0 40 0 40 0 30 0 20 0 20 0 30

ERKpp (IF)

0 10

n = 53

n = 159

n = 168

n = 156

n = 170

n = 159

n = 267

n = 55

0 0.1 0.2 0.5 1 2 5 10 PDGF

PDGF (ng·mL–1) 200

400 600 800

Hill coefficient 3.8 ± 0.7

0 20 40 60 80

100

A

B

C

mRNA half-life (h)

Fig 4 Validation of model predictions (A) Cumulative distribution

of mRNA half-lives DUSP6 has a very low half-life (median 0.55 h, marked with an arrow), which is significantly smaller than average mRNA half-lives (B, C) Distribution of ERK phosphorylation is single cells after platelet-derived growth factor (PDGF) stimulation shows that ERK responds with a unimodal distribution (B), in an ultrasensi-tive fashion at the population level with a Hill coefficient of about

4 (C).

DUSP6 protein and mRNA, respectively Recently,

two studies of mRNA decay rates, measured at the

level of the genome, showed a median half-life of

DUSP6 mRNA of 33 min This is one of the shortest

half-lives in the entire dataset (Fig 4A) [35,36] In

addition, the half-life of the DUSP6 protein has been

reported to be < 1 h [37] Thus, the model prediction

of very short half-lives for DUSP6 is congruent with

the data available

Model prediction 2 – ERK activity is ultrasensitive

The next prediction of the model selection procedure is

that ERK activation is ultrasensitive Possible

mecha-nisms underlying ultrasensitivity have been discussed

earlier The most likely mechanism is distributed,

sequential phosphorylation⁄ dephosphorylation of

ERK, which typically gives rise to a Hill coefficient of

2 [33] It was not possible to reliably measure small

quantitative changes in MEK activation in our

experi-mental system Therefore, we tested this prediction by

stimulating fibroblasts with different concentrations of

the platelet-derived growth factor (PDGF), and

mea-sured ERK phosphorylation in the nucleus 20 min

poststimulation by immunofluorescence Single-cell

measurements were employed, as it has been proposed

earlier that PDGF-stimulated fibroblasts react in a

bistable manner, with individual cells responding in an

all-or-none fashion [8,38] Such cellular behavior is

expected to give rise to a bimodal histogram of ERK

activity for intermediate stimuli The distribution of

ERK activity is shown in Fig 4B In contrast to

bista-ble responses, the stimulation experiments showed a

monomodal distribution of ERK activity, which

grad-ually shifts to higher activity levels as the stimulus

increases This suggests that ERK activity is not

bista-ble in fibroblasts stimulated with PDGF The average

activity shows a Hill-type response with a coefficient of

approximately 4 (Fig 4C) Thus, ERK activation is

ultrasensitive when it responds to PDGF stimulation

As Hill coefficients are determined in cell populations,

the strong sensitivity of the response observed at the

population level could be even more pronounced at

the level of individual cells A similar estimate for the

Hill coefficient can be derived from previously

pub-lished data obtained by flow cytometry [39] (for

details, see Appendix S3) Such ultrasensitivity may

arise at any point during the transduction from

recep-tor to ERK Ultrasensitivity is partly due to the

func-tion of the receptor, which has been determined to

respond with a Hill coefficient of 1.7 in fibroblasts

[40] Hill coefficients of signaling cascades are

maxi-mally the product of the Hill coefficients of the

individual elements of the signaling pathway [41,42] Therefore, the remaining coefficient of at least 2 can

be attributed to MAPK signaling It remains to be shown, however, whether the resulting ultrasensitivity results from the double phosphorylation of ERK, or from a combination with processes further upstream, such as MEK phosphorylation

Discussion MAPK signaling is central to proliferation control in many cells, and quantitative aspects of ERK activa-tion, such as signal amplitude and duraactiva-tion, determine the cell fate However, little is known of how MAPK signaling is regulated quantitatively by transcriptional feedback loops, although the time scale of decision-making is often well beyond that of early transcrip-tional feedbacks To improve our knowledge of the transcriptional responses involved in MAPK signaling,

we have employed a systems biological approach to identify a feedback loop that shapes the activation of ERK within the first hours of cellular transformation

We present evidence that DUSP6 is transcriptionally upregulated by oncogenic RAS signaling through the potential cooperation of SRF with CREB, and thus causes a biphasic response of ERK Current sequence-based methods fail to provide a genome-wide predic-tion of target genes, due to the high number and length of the mammalian promoters and the short binding motifs of transcription factors However, the combination of ‘conventional’ promoter analysis with gene ontology-term-based functional annotation [22] revealed phosphatase genes as primary targets of SRF and of CREB SRF is a key determinant of muscle dif-ferentiation, and plays a major role in the regulation

of proliferation through the activation and repression

of a variety of target genes [43,44] The transcriptional activity of SRF is stimulated through ERK-dependent phosphorylation Specificity is achieved by interaction

of SRF with cofactors in a signal-specific or tissue-specific manner These cofactors bind either together with SRF at the serum response element or in close proximity to Ets binding sites [45] Also, the CREB transcription factor has been implicated in the regula-tion of proliferaregula-tion, mainly in leukemias through the induction of proto-oncogenes and cell cycle regulatory genes [46] A direct impact of CREB-mediated gene activation on signaling feedback control during trans-formation or tumor development has not been reported previously Thus, our approach identified a hitherto unknown combinatorial role of both tran-scription factors, which is likely to determine both the onset and quantity of mitogenic signaling in several

different cellular contexts In a recent publication [44],

DUSP6 was predicted to harbor an SRF-binding site;

however, this was not confirmed experimentally

Therefore, it remains to be tested whether CREB and

SRF both interact with the phosphatase genes, or

whether there is an indirect contribution of SRF,

which does not seem likely, because of the rapid

induc-tion SRF might also play a role as a mediator for

MAPK-dependent, ETS1-controlled induction of

DUSP6, as suggested very recently [47] The fact that

we found phosphatases to be overrepresented in the

joint list of SRF and CREB targets suggests that

regulation of protein phosphorylation is a common

function of the two transcription factors

Another important aspect of our model is that it

predicted the very short half-lives of DUSP6 mRNA

and protein, which have been reported to be £ 1 h

[37] The time span required for a protein to reach a

steady-state expression level is determined by its

half-life [48] Therefore, short-lived molecules such as

DUSP6 can respond quickly to any alteration in

signaling, and thus can influence ERK activity within

1–2 h The functional relevance of DUSP-mediated

feedback is supported by a recent study using

meta-bolic control analyses on epidermal growth factor

receptor models [49] This study predicted a central

role for the dephosphorylation of ERK The

distribu-tion of control strength within the epidermal growth

factor receptor-induced network of MAPK signaling

showed that relatively few, distinct steps in the

signal-ing cascade appeared to have a significant control

function for the signaling amplitude, duration and

integrated output of transient ERK phosphorylation

The dephosphorylation of ERK by DUSP6 and also

the overall protein concentrations of both ERK and

DUSP6 had a significant influence on signaling control

[49] Such differential control functions might have

important implications for the efficacy of targeted

pathway inhibition, as blocking of different pathway

components might cause different and eventually

unex-pected biological responses One important aspect is

that systems controlled by negative feedbacks may be

very robust with respect to manipulations of different

components within the feedback loop [50] MAPK

phosphatase genes, such as DUSP6 and DUSP4, are

at least partially understood in terms of their

transcrip-tional regulation downstream of ERK [44,47,51] In

addition, there are other feedback regulators within

the RAS–RAF–MEK–ERK pathway that might

con-tribute to signaling modulation We tested whether

feedback regulation via DUSP1, an unspecific MAPK

phosphatase, contributes to early signal attenuation,

but found that it does not play a major role, possibly

because the cells are not serum-starved DUSP9, which

is induced at later a time point, may mediate signal attenuation at time points after 10 h Other classes of signaling proteins might also mediate transcriptional feedback Most recently, Ding and Lengyel [52] described a novel regulator of RAS, p204, which is induced by Egr1, a transcription factor directly down-stream of ERK Moreover, Sprouty, an inhibitor act-ing at the receptor level, seems to be transcriptionally regulated upon pathway activation [48] Therefore, fur-ther quantification and more detailed modeling includ-ing time-resolved analysis of phosphatase expression will be required to determine whether therapeutic approaches targeting DUPSs or signaling components further downstream of MAPK could be beneficial Several lines of evidence suggest that the feedback mediated by DUSP6 ‘steps in’ whenever noncancerous cells are exposed to prolonged stimulation This allows switching-off of the pathway [53] Several studies have demonstrated the role of DUSP6 as a central feedback regulator dampening ERK levels in developmental programs [14,54] Our study shows that a strong onco-genic signal can overcome this negative feedback and achieve constitutive ERK activation However, it also shows that the feedback keeps ERK activity at a mod-erate level One could speculate that the robustness gained from this feedback in normal cells is ‘hijacked’

or co-opted by cancer cells to circumvent apoptosis caused by ERK overactivation The role of DUSP6 in controlling the robustness of tumor cell proliferation and progression seems to be dependent on tumor type Pancreatic cancer cells progress towards a more aggressive and invasive phenotype following loss of DUSP6 expression [55] In breast cancer cells, activa-tion of the DUSP6 feedback correlated with chemo-therapy resistance following tamoxifen treatment [56] Moreover, DUSP6 is part of a predictive gene signa-ture for non-small cell lung cancer based on five infor-mative genes [57] These examples show that it is crucial to understand MAKP-dependent control mech-anisms in more quantitative terms, and suggest that molecules involved in feedback regulation can play ambiguous roles as oncogenes and tumor suppressors, depending on quantitative differences

From our experimental data, we could derive a role for one of the regulated DUSPs As other DUSPs are regulated as well, MAPK signaling is most likely lated by a complex network of negative feedback regu-lators Moreover, the stability of several DUSP proteins is regulated by post-translational modification [37] Our study based on model selection and time course data could not fully resolve the complexity of this regulatory network downstream of MEK In order

to disentangle this network, a much more complex

study needs to be conducted, including pathway

inter-ference, and incooperating biophysical data such as

binding constants and protein concentrations, which

crucially influence the dynamics of the pathway [33]

Only then we will be able to understand why such a

complex network of negative feedback players controls

MAPK signaling

Moreover, the biological variability in our

experi-mental system, which caused different induction

kinet-ics of RAS, was a limitation, as all data used to

calibrate the model had to come from one

experimen-tal time course In future studies, other means of

receptor-independent stimulation need to be exploited

Our study also warrants the conclusion that

mathe-matical modeling of signaling pathways needs to

incor-porate the response of the transcriptome, if it is aimed

at modeling the pathways for physiologically relevant

time intervals Previous detailed mathematical models

have emphasized the importance of post-translational

feedbacks, but have generally neglected transcriptional

feedback loops A recent analysis has shown that

tran-scriptional feedback regulation by short-lived

inhibi-tory molecules controls all major signaling pathways in

humans [48] Therefore, we expect that similar

semi-quantitative studies on the feedback regulation of

other disease-related pathways are required to fully

appreciate the complexity of pathway control Such

studies could guide searches for new and more

patient-tailored therapeutic interventions and provide solutions

that either bypass the feedback loops or even modulate

the loops and achieve high therapeutic potential

Experimental procedures

Cell culture conditions, transfection and

imunofluorescence

derivatives FE-8 [58] and NIH3T3 cells were cultured in

DMEM supplemented with 10% fetal bovine serum, 2%

-transformed human embryonal kidney cells were described

by Hahn et al [26], and were cultivated in MEM, alpha

modification, supplemented with 10% inactivated fetal

Transient siRNA transfections against SRF were

per-formed for 96 h after double transfection with two different

oligonucleotides: SRF-1 (UGAGUGCCACUGGCUUUG

Att sense, UCAAAGCCAGUGGCACUCAtt antisense),

constructed with the Silencer siRNA Construction Kit (#1620; Ambion, Applied Biosystems, Carlsbad, CA, USA) and SRF-2 predesigned by Ambion (ID 142734) In both cases, a final concentration of 50 nm was used

Immortal rat 208F fibroblast-derived inducible RAS (IR) cells (clone IR-4) harbor an IPTG-inducible HRAS onco-gene, and have been described previously [29] Expression

of HRAS was induced by the addition of 20 mm IPTG NIH3T3 cells grown on coverslips were serum-starved for

48 h and then treated with increasing concentrations of PDGF PhosphoERK immunofluorescence was determined

USA) after 15 min of fixation in 3%

pMAPK antibody for 2 h and with an Alexa546-labelled antibody against rabbit for 1 h Pictures were taken with a

described in Appendix S3

Western blot analyses

pH 8.0, 100 mm NaCl, 1% sodium deoxycholate, 1% NP-40, 0.1% SDS, complete protease inhibitor mix (Roche, Mannheim, Germany)], and 20 lg of the whole cell extracts

(TransBlot SD; BioRad, Laboratories, Munich, Germany)

to polyvinylidenefluoride membranes (Hybond P; Amer-sham, Little Chalfont, UK), the membranes were blocked

0.05% Tween-20) with 5% nonfat dry milk, and incubated with primary antibodies against RAS (Transduction Lab-oratories, BD Biosciences, San Jose, CA, USA) and

Biolabs) Membranes were washed and incubated with

detected by chemiluminescence reaction (ECL; Amersham Pharmacia, Little Chalfont, UK) according to the manufac-turer’s instructions

Microarray experiments Predesigned 70-mer oligonucleotides produced by Illumina Inc (San Diego, CA, USA) were spotted at 20 lm in 3· SSC buffer, containing 0.01% SDS, onto poly(l-lysine)-treated glass slides Spotting was performed with the Micro-Grid microarrayer (Genomic Solutions, Ann Arbor, MI, USA) Every oligonucleotide was spotted six times In addi-tion, 20 different housekeeping genes and positive and negative controls provided by the Alien SpotReport cDNA Array Validation System were included (Stratagene, La Jolla, CA, USA) Labeling and microarray hybridization

was performed manually according to the Genisphere

3DNA Array 50 kit protocol (Genisphere, Hatfield, PA,

experi-ment was performed

Microarrays were scanned with two wavelengths for Cy3

(570 nm) and Cy5 (660 nm) by using a laser fluorescent

scanner (Agilent G2565BA Scanner; Agilent Technologies,

Palo Alto, CA, USA) with three different photomultiplier

version 3.0 (BioDiscovery, Los Angeles, CA, USA) Raw

data obtained with the highest photomultiplier gain were

routinely used for quantification Spots with saturated

sig-nal intensity were reasig-nalyzed using a lower photomultiplier

gain The fluorescence intensity of each spot in both the

Cy3 and Cy5 images was quantified, and fluorescence levels

of the local background were subtracted Normalization of

Cy3 and Cy5 images was performed by adjusting the total

signal intensities of two images A Lowess curve was fitted

to the log intensity versus log ratio plot Twenty per cent of

the data were used to calculate the Lowess fit at each point

This curve was used to adjust the control value for each

measurement If the control channel was lower than 10,

then 10 was used instead

Northern blot analysis

The RNA was transferred to a nylon membrane (Nytran N;

Schleicher & Schuell, Dassel, Germany) and crosslinked by

in hybridization buffer (ExpressHyb; Clontech, Takara

yeast tRNA Twenty-five nanograms of the cDNA probe

c.p.m of the labeled probe was added per milliter of hybridization buffer and

a stringency of 2· SSC ⁄ 0.1% SDS at 42 C, exposed to

equal loading and integrity of RNA, all gels were stained

with ethidium bromide mRNA levels were normalized with

glyceraldehyde 3-phosphate dehydrogenase or 18S rRNA

Real-time PCR analysis

RNA was prepared as described above 96 h after the

sec-ond siRNA transfection Expression patterns of the genes

were validated by real-time RT-PCR using the ABI Prism

7900HT Sequence Detection System and TaqMan Gene

Expression Assays (Applied Biosystems, Foster City, CA,

USA), according to the supplier’s instructions For relative

quantification, the linear expression values were calculated

by the DDCT method [59], using the tubulin gene as an

internal control

Acknowledgements

We thank Dr Thomas Korte and Professor Andreas Herrmann, Institute for Biophysics, HU Berlin for help and advice on fluorescence microscopy This pro-ject was funded by Deutsche Forschungsgemeinschaft DFG, SFB 618 Theoretische Biologie, projects A1 and A3 and by the German Ministry for Education and Research (BMBF), through the FORSYS partner programme (grant number 0315261)

References

1 Murphy LO & Blenis J (2006) MAPK signal specificity: the right place at the right time Trends Biochem Sci 31, 268–275

2 Klein PJ, Schmidt CM, Wiesenauer CA, Choi JN & Gage EA (2006) The effects of a novel MEK inhibitor PD184161 on MEK–ERK signaling and growth in human liver cancer Neoplasia 8, 1–8

3 Bluthgen N & Legewie S (2008) Systems analysis of MAPK signal transduction Essays Biochem 45, 95– 108

4 Orton RJ, Sturm OE, Vyshemirsky V, Calder M, Gil-bert DR & Kolch W (2005) Computational modelling

of the receptor-tyrosine-kinase-activated MAPK path-way Biochem J 392, 249–261

5 Vayttaden SJ, Ajay SM & Bhalla US (2004) A spectrum

of models of signaling pathways Chembiochem 5, 1365– 1374

6 Huang CY & Ferrell JE Jr (1996) Ultrasensitivity in the mitogen-activated protein kinase cascade Proc Natl Acad Sci USA 93, 10078–10083

7 Santos SD, Verveer PJ & Bastiaens PI (2007) Growth factor-induced MAPK network topology shapes Erk response determining PC-12 cell fate Nat Cell Biol 9, 324–330

8 Bhalla US & Iyengar R (1999) Emergent properties of networks of biological signaling pathways Science 283, 381–387

9 Asthagiri AR & Lauffenburger DA (2001) A computa-tional study of feedback effects on signal dynamics in a mitogen-activated protein kinase (MAPK) pathway model Biotechnol Prog 17, 227–239

10 Kholodenko BN (2000) Negative feedback and ultrasen-sitivity can bring about oscillations in the mitogen-acti-vated protein kinase cascades Eur J Biochem 267, 1583–1588

11 Nakayama K, Satoh T, Igari A, Kageyama R & Nish-ida E (2008) FGF induces oscillations of Hes1

R332–R334

12 Eblaghie MC, Lunn JS, Dickinson RJ, Munsterberg

AE, Sanz-Ezquerro JJ, Farrell ER, Mathers J, Keyse

SM, Storey K & Tickle C (2003) Negative feedback