Investigating Differential Dynamics of the MAPK Signaling Cascade Using a Multi-Parametric Global Sensitivity Analysis Authors: Jeongah Yoon, Thomas S. Deisboeck Published: 2009, PLOS ONE DOI: 10.1371/journal.pone.0004560 ABSTRACT Cell growth critically depends on signalling pathways whose regulation is the focus of intense research. Without utilizing a priori knowledge of the relative importance of pathway components, we have applied in silico computational methods to the EGF-induced MAPK cascade. Specifically, we systematically perturbed the entire parameter space, including initial conditions, using a Monte Carlo approach, and investigate which protein components or kinetic reaction steps contribute to the differentiation of ERK responses. The model, based on previous work by Brightman and Fell (2000), is composed of 28 reactions, 27 protein molecules, and 48 parameters from both mass action and Michaelis-Menten kinetics. Our multi-parametric systems analysis confirms that Raf inactivation is one of the key steps regulating ERK responses to be either transient or sustained. Furthermore, the results of amplitude-differential ERK phosphorylations within the transient case are mainly attributed to the balance between activation and inactivation of Ras while duration-differential ERK responses for the sustained case are, in addition to Ras, markedly affected by dephospho-/phosphorylation of both MEK and ERK. Our sub-module perturbations showed that MEK and ERK's contribution to this differential ERK activation originates from fluctuations in intermediate pathway module components such as Ras and Raf, implicating a cooperative regulatory mode among the key components. The initial protein concentrations of corresponding reactions such as Ras, GAP, and Raf also influence the distinct signalling outputs of ERK activation. We then compare these results with those obtained from a single-parametric perturbation approach using an overall state sensitivity (OSS) analysis. The OSS findings indicate a more pronounced role of ERK's inhibitory feedback effect on catalysing the dissociation of the SOS complex. Both approaches reveal the presence of multiple specific reactions involved in the distinct dynamics of ERK responses and the cell fate decisions they trigger. This work adds a mechanistic insight of the contribution of key pathway components, thus may support the identification of biomarkers for pharmaceutical drug discovery processes. INTRODUCTION Deregulated cell growth, a hallmark of cancer, is associated with perturbed signal transduction [1] . In response to external stimuli by specific ligands, receptor tyrosine kinases (RTKs) can alter cellular phenotypes such as cell survival, proliferation and migration [2] . There have been a number of different intracellular signaling pathways activated by RTKs, including epidermal growth factor (EGF) signaling [3] . The output of the signal transduction frequently targets the activation of the extracellular signal-regulated kinase (ERK), which is known to be involved in solid tumor formation by regulating cell cycle progression [4] , [5] . Previous experimental studies support differential duration of ERK activity as being critical for cell signaling decisions [6] – [11] . However, it still remains uncertain which mechanisms control those phenotypic decisions. Thus, further characterizing the sub-cellular reaction steps or protein molecules that are critical for differential control of cellular activities is important not only to better understand the dynamics of complex signaling system for predictive purposes, but also to identify potential therapeutic targets for drug development [12] . The continuous advancement of high-throughout technologies in the post genomic era presents the challenge of how to interpret an ever growing amount of molecular data. Numerous experimental works have attempted to identify particular signaling molecules and their mechanisms, for example, by constructing mutants, overexpression, or reconstitutions of arrangement of genes or proteins [6] . Using experimental techniques alone, however, will make it difficult to identify decisive reaction steps or molecules that control for instance differential ERK responses. This is not only due to the inherently complex structure and function of various signal transduction pathways, but also because of marked nonlinearity within the system. Thus, it is useful to introduce computational, data- or hypothesis-driven approaches in an effort to facilitate the discovery of cell signaling decision factors [13] . Also, complex signaling pathways can be treated as a tightly connected network that orchestrates regulation of a specific functional behavior, rather than as individual, separate mechanisms by focusing on a specific oncogenic molecule or its activating event alone [14] . In this paper, we therefore apply a systems-level, multi-parametric perturbation strategy using a Monte Carlo (MC) simulation to discover molecules or reaction steps that orchestrate differential mitogen-activated protein kinase (MAPK) signaling responses. The model system is an EGF-induced signaling pathway, originally compiled by Brightman and Fell [15] . We have disturbed every single parameter without a priori knowledge on the relative importance of certain parameters and have generated massive samples of multiple perturbations for all parameters using our MC simulation; the peak amplitude (height) and duration (width) of ERK profiles (e.g., transient vs. sustained ) are then used as differentiation measures [16] , [17] . Our analysis reveals the dominant role of intermediate module proteins such as Ras and Raf as key controlling factors for the distinct dynamics of ERK activation. Although MEK and ERK in the MAPK module also showed sensitivities, they alone did not affect differential ERK dynamics without the co-perturbation of intermediate module proteins. This may implicate a cooperative regulation mode of key components in cellular responses. In addition, initial concentrations of the key proteins in corresponding reactions are also actively involved in determining the cellular responses. Lastly, we note that here identified critical molecules have already been experimentally validated as biomarkers. METHODS The EGF receptor (EGFR) system implemented is based on the Brightman and Fell model ( Figure 1 ) [15] . Although this pathway representation is moderately-sized, it covers the major cascade of an EGF-induced Ras-dependent MAP kinase signaling pathway with one feedback loop. Here, we provide a brief biological description of the signaling cascade. This network is divided into three subsystems according to somewhat separable roles and topographic locations. Top-level, intermediate, and MAP kinase modules are colored in orange , magenta , and navy , respectively. Filled arrowheads represent protein-protein binding reactions whereas open arrowheads denote catalytic interactions. Rs, free cell surface receptor; L, ligand; Ri, internal receptor; Li, internal ligand; RL, cell surface R-L complex; R2L2s, cell surface dimmer complex of RL; R2L2i, internalized dimmer complex of RL; CPP, coated pit protein; R2L2-CPP, receptor-ligand-coated pit protein complex; R*, all active species of R, i.e., R2L2s, R2L2i, and R2L2-CPP; Shc, Src homology and collagen domain protein; ShcP, phosphorylated Shc; GS, complex of Grb2, growth factor receptor binding protein 2, and SOS, Son of sevenless homologue protein; GSP, phosphorylated GS; ShcGS, complex form of Shc and GS; Ras-GDP, inactive, GDP-bound Ras; Ras-ShcGS, complex form of Ras-GDP and ShcGS; Ras-GTP, active, GTP-bound Ras; GAP, Ras GTPase activating protein; MEKP and MEKPP, phosphorylated forms of MEK; ERKP and ERKPP, phosphorylated forms of ERK. The mechanisms in the first or top module occur close to the plasma membrane, and are related to the activation of the EGF receptor. First, the corresponding ligand EGF (L) as the only external stimulus binds to a monomeric receptor (R), forming an RL complex prior to being dimerized; at this point, intrinsic protein tyrosine kinases are activated. Only the activated dimer complex species are internalized through binding to cell-surface coated pit adaptor proteins (CPP). These complexes are then dissociated and degraded, and the monomeric species are recycled to the plasma membrane. In the second or intermediate module, the activated receptor (R*) catalyzes the adaptor protein Src homology and collagen domain protein (Shc). Phosphorylated Shc (ShcP) then forms a ternary complex, ShcGS, with a constitutive complex of growth factor receptor binding protein 2 (Grb2) and Son of sevenless homologue protein (SOS), the guanylnucleotide exchange factor. Subsequently, the ShcGS complex recruits cytoplasmic SOS to the membrane-bound Ras protein, where the inactive Ras-GDP is activated to Ras-GTP through interaction with ShcGS. There are two downstream options with regards to active Ras-GTP: it either binds to GAP (GTPase-activating protein) to stimulate the GTPase activity of Ras so that RasGTP is converted to inactive RasGDP; or, Ras-GTP binds to Raf, forming the Ras-Raf complex such that Raf is recruited to the plasma membrane before the complex facilitates the Raf kinase activation (Raf*). The activated Raf in the latter option phosphorylates the mitogen-activated protein kinase (MAPK) cascade, which constitutes the third module. In this third or MAPK module, both MEKP and MEKPP activate ERK by phosphorylating a tyrosine and a threonine residue, but only ERKPP is known to be active. ERKPP leads to the dissociation of the ShcGS complex through feedback regulation [18] . Raf*, MEKP, MEKPP, and ERKPP are dephosphorylated by the same phosphatase, protein phosphatase 2A (PP2A). (We note that there have been several studies that focused on understanding the dynamics of phospho-/dephosphorylation of this MAPK cascade module by combining in silico modeling with experimental data [19] – [23] ). MULTI-PARAMETRIC GLOBAL SENSITIVITY ANALYSIS The EGFR signaling cascade system implemented here consists of 28 kinetic reactions involving 27 different protein molecules and 48 parameters. In general, these reactions follow mass action kinetics except for those catalyzed by enzymes, which follow Michaelis-Menten (MM) kinetics. The main goal of procedures detailed in this section is to calibrate key pathway elements (e.g., parameters, molecules, or reaction structures) that are chiefly responsible for processing cellular phenotypic decisions within a tolerable range. To this end, any approaches may face the following two challenges: First, even for a moderately sized cell signaling network, it is non-trivial to track the molecules' temporal behavior as kinetic parameters are often still unknown and difficult to measure experimentally. Second, even for known parameters, detailed quantitative measurements of protein activities may have been conducted under experimental condition that are far from the realities of an in vivo environment and that vary between experiments [24] . Thus, one may have to rely on potentially inaccurate measurements of input parameter sets. We have therefore applied an approach that can avoid heavy dependence on a given set of initial parameter values. At the core is a Monte Carlo (MC) simulation that explores certain ranges of the parametric space around a given initial parameter value and generates samples of numerous parameter vectors. These vectors contain a set of multiply perturbed individual elements and are randomly generated using uniform distribution functions within known ranges of parameter uncertainty. This MC simulation is used for the entire parametric uncertainty analysis (see Figure S1 ). The detailed procedure for our MC simulation-based global uncertainty analysis is as follows [25] , [26] : STEP 1 Define a range for k parameters (or every initial condition) involved in the signaling cascade, reflecting the uncertainties of signaling responses. The lower and upper bound of ranges has been suggested by [27] , which reflects a variation of approximately (+/−) 2 orders of magnitude for the initial value of each parameter. STEP 2 Generate a series of independent random numbers using a uniform distribution for each parameter within defined ranges of uncertainties at Step 1. The total number of generated samples (N = 10,000) is assumed to be independent of each other and also sufficiently large in number. STEP 3 Run the ordinary differential equation (ODE) model for each set of k parameters and calculate an objective function value for the ERK profile. The objective function is defined as the sum of the squared errors of the active ERK level between the unperturbed ( unper ) and the perturbed ( per ) system as follows: (1) where j is the number of time points and i is the total number of samples generated by the MC simulation ( i  = 1, …, N, N = 10,000). STEP 4 Compare the objective function value to a threshold value. In this study, the threshold is defined as the sum of the squared errors between the active ERK profile from the unperturbed system and the average active ERK profile from all samples. Based on the threshold, each parameter set is classified into either a tolerable sample group ( m samples) when the error sum of the ERK signal from a certain parameter set is below the threshold, or an intolerable sample group ( n samples) when it is above the threshold. Using the threshold, 99% of the generated samples have been classified into the tolerable group. We only retain the tolerable group samples and discard the others. Note that the sum of m and n is equal to N, the total number of samples generated by the MC method. STEP 5 Distinguish differential profiles of ERK responses using tolerable group samples only. In this study, we consider three cases of two possible differential ERK responses: i) transient ERK level (T) vs. sustained level (S), ii) lowly transient ERK level (L-T) vs. highly transient level (H-T), and iii) lowly sustained ERK level (L-S) vs. highly sustained level (H-S). In order to classify samples of the tolerable group into the two types for each case, we introduce two characteristic measures, i.e., amplitude and duration of the ERK profile. In this study, we define the ‘amplitude’ as the maximum level of ERK over a time period of 60 min and the ‘duration’ as the time period from the point of the maximum ERK level to the point of reaching 10% of the maximum, within 60 min ( Figure 2 ). a) definitions of amplitude and duration (the ‘amplitude’ is defined as the maximum level of ERK over a time period of 60 min, and the ‘duration’ as the time period from the point of the maximum ERK level to the point of reaching 10% of the maximum within 60 min), b) transient ERK level (T) vs. sustained level (S), c) lowly transient ERK level (L-T) vs. highly transient level (H-T), and d) lowly sustained ERK level (L-S) vs. highly sustained level (H-S). In order to efficiently classify and collect samples from the tolerable sample group for each case, we first sorted the samples with the maximum amplitude of ERK in ascending order. Then, for case 1 (T vs. S), transient samples are collected as those satisfying the criterion that the ERK level at the last time-point observation (i.e., at 60 min) is less than 10% of the maximum amplitude; sustained samples are collected according to the maximum duration, in addition to considering the maximum amplitude. For case 2 (L-T vs. H-T), L-T group samples are those below the median profile of ERK in case 1; H-T samples are those above the median. For case 3 (L-S vs. H-S), we further extracted samples with the duration of more than 30 min from the sorted samples with the maximum amplitude level in case 1. Because the maximum amplitude of ERK often occurs within the first 10 to 20 min (within the 60 min period), we assumed sustained samples would have the duration of more than 30 min; accordingly, samples of the duration of less than 30 min have been discarded. From the extracted sample list (ordered from the sample with the longest duration to that with the shortest) we have collected L-S samples from the bottom (shortest duration sample) of the list, while H-S samples have been taken from the top (longest duration sample) of the sample list. Selected were 367 samples for T and 500 samples for S in case 1, 365 samples for L-T and 367 samples for H-T in case 2, and 100 samples for both the L-S and H-S in case 3. Note that the number of samples for each group is arbitrarily chosen. During the process, our goal was that collected samples for each case have distinctively separable characteristics, so that results from the multi-parametric global sensitivity analysis can provide recognizable features for each comparison. STEP 6 Evaluate parametric sensitivities by comparing the parameter distributions between two sample sets of differential ERK responses for all three cases (i.e., T vs. S, L-T vs. H-T, and L-S vs. H-S). Here, we have simply calculated cumulative frequency (CF) distributions to identify informative parameters and reactions that contribute to the difference between two differential responses. For instance, if the CF distributions between the two groups for a certain parameter are distinctively different, i.e., yielding low correlation coefficients between the two CF distributions, the parameter is classified as a sensitive, fragile, or informative factor because it contributes to the control of a particular type of differential ERK responses; otherwise, it is classified as an insensitive, robust, or uninformative factor. In the following section, we briefly introduce overall state sensitivities (OSSs) [28] . OSS is obtained by perturbing one single parameter at a time while keeping all other parameters fixed. This is in contrast to our MC-based approach that changes multiple parameters simultaneously. We will then analyze how the single parametric global sensitivity analysis can strengthen or weaken the results from the multi-parametric approach. OVERALL STATE SENSITIVITY ANALYSIS The OSS index is often used to capture global robustness of state variables upon parameter perturbations [29] . For example, in a simple enzyme mediated reaction, let the Michaelis constants Km be a parameter p , and the concentration of the substrate or activation level of the protein be a variable X . Then the parameter sensitivity of X regarding p is defined by (2) With this general definition in mind, to calculate the overall state sensitivity for the individual element j of a perturbed parameter set, the OSSs are integrated over discrete time as described previously [28] : (3) Here, N T represents the number of time points while N S denotes the number of protein molecules or protein complexes in the system. OSS describes how robust a system is to a single parameter change while the other parameters are fixed. We perturbed at most (+/−) 50% of the original parameter value. Note that all parameter sensitivities are only valid in a local space, i.e., within the proximal space of the unperturbed parametric space. All numerical simulations of biochemical reaction-ODEs and MC-based simulations were implemented in MATLAB (version 7.5.0, MathWorks, Natick, MA). We used ode15s function for solving the nonlinear ordinary differential equations with 300 time points for a 60 min simulation. Using a workstation with a 2.00 GHz CPU and 4.00 GB of RAM it took approximately 3.5 hours of CPU time to simulate 10,000 samples for the entire pathway run. RESULTS The nominal parameter set, initial conditions, and their perturbed ranges are depicted in Table 1 and Table 2 . Note that the nominal parameter set of the EGFR model gives a transient pattern of ERK activity. In this section, we show results of comparing the aforementioned three cases: i) transient vs. sustained ERK profiles, ii) lowly transient vs. highly transient ERK profiles, and iii) lowly sustained vs. highly sustained ERK profiles. For each case, we examine which pathway components influence the differential responses most. The column of actual test range represents parameter ranges that are contained in the generated samples. The column of actual test range represents ranges of initial values that are contained in the generated samples. First, we investigated the differentiating pathway parameters between transient and sustained activation of ERK profiles. The cumulative frequency (CF) distributions of transient and sustained cases for each parameter are shown in Figure 3 ; correlation coefficients between the two CF distributions are shown in Table S1 , where larger correlation coefficients represent more robust parameters. Solid lines in red and in blue represent the S and T case, respectively. The multi-parametric analysis based on the MC method identifies k14, k16, k19, and k25 as the most sensitive forward mass-action kinetic parameters while no significant effects result from reverse kinetic parameters. Furthermore, the reaction rates of V18, V20, and mildly V24 and V26 appear to be closely involved in controlling the differential ERK responses. Finally, with regards to the Michaelis constants, Km18 and Km22–Km26 are the dominating parameters. Together, this indicates that the reaction steps R14, R16, R18, R19–R20, R22, R24, and R26 are those primarily controlling differential ERK responses (i.e., a transient vs. a sustained activation profile). The corresponding protein molecules involved in those reactions include active Ras and Raf (RasGTP and Raf*), MEK, and ERK. In taking a detailed look at the reactions, first, R14 and R16 are those that balance active Ras between an inactivated and activated state either by binding to GAP (to be finally converted to its inactive form (RasGDP)) or by binding to Raf (to further process the downstream pathway of active Raf). This activation of Raf becomes important for initiating the phosphorylation of MEK to MEKP through R19 (Note: reaction steps from R19 to R26 denote the phosphor-/dephosphorylation reactions in the MAPK module). The reverse reaction (R18) that dephosphorylates active Raf (Raf*) into its inactive form is also found to be a sensitive, controlling factor. We also find that those samples that show transient ERK activation tend to have smaller values of k14, V20, Km18, Km22 (and less so of Km25), but larger values in k16, k19, V18, and Km 24, Km26 (see Figure S2 ). For instance, the Ras dephosphorylation reaction rate (k14) is slower for the transient case than for the sustained one. Besides, higher frequencies of smaller Km18 values indicate that the affinity of active Raf to its specific enzyme PP2A is far more stable and stronger, resulting in a much faster Raf inactivation in the transient as compared to the sustained case. The higher frequency distribution of smaller Km22 also indicates faster MEKPP dephosphorylation in the transient case ( Figure 4 ). Similarly, the larger k16 value expresses faster reaction activity in the transient case, and the distribution of larger values of parameter V18 indicates that R18 occurs much faster as well ( Fig. 4 ). Lastly, larger constants Km23, Km24, and Km26 imply less active ERK phospho-/dephosphorylation reactions in the transient case, compared to the sustained case. These observations reveal that although the downstream signal to the ERK activation occurs much faster, it is ‘short-lasting’, thus producing a transient behavior. In contrast, for the sustained ERK response, active Raf (Raf*) lasts longer so that dephosphorylated MEKP and MEKPP have a higher chance to be re-activated, which will eventually produce a prolonged signal for ERK activation. Taken together, the association of both active Ras and Raf are crucial for controlling the response of ERK in the MAPK signaling cascade. Furthermore, reactions related to kinases and phophatases for both MEK and ERK are also shown to be important for distinguishing between transient or prolonged ERK activation as a consequence of EGF stimulation. The cumulative frequency (CF) distribution informs that the higher frequencies of larger Km18 and smaller V18 in the sustained case indicate a much slower reaction rate of R18 in the sustained case than in the transient case (i.e., v s S) Sustained dominant (TH-T) Highly-transient dominant (L-TH-S) Highly-sustained dominant (L-S