Abstract
Background
The tumor suppressor protein p53 plays important roles in DNA damage repair, cell cycle arrest and apoptosis. Due to its critical functions, the level of p53 is tightly regulated by a negative feedback mechanism to increase its tolerance towards fluctuations and disturbances. Interestingly, the p53 level is controlled by posttranslational regulation rather than transcriptional regulation in this feedback mechanism.
Results
We analyzed the dynamics of this feedback to understand whether posttranslational regulation provides any advantages over transcriptional regulation in regard to disturbance rejection. When a disturbance happens, even though negative feedback reduces the steadystate error, it can cause a system to become less stable and transiently overshoots, which may erroneously trigger downstream reactions. Therefore, the system needs to balance the tradeoff between steadystate and transient errors. Feedback control and adaptive estimation theories revealed that posttranslational regulation achieves a better tradeoff than transcriptional regulation, contributing to a more steady level of p53 under the influence of noise and disturbances. Furthermore, posttranslational regulation enables cells to respond more promptly to stress conditions with consistent amplitude. However, for better disturbance rejection, the p53 Mdm2 negative feedback has to pay a price of higher stochastic noise.
Conclusions
Our analyses suggest that the p53Mdm2 feedback favors regulatory mechanisms that provide the optimal tradeoffs for dynamic control.
Keywords:
Feedback control theory; p53Mdm2 feedback loop; Robustness; Disturbance rejectionBackground
Gene networks are constantly subject to noise or fluctuations, which originate from variations in transcription, translation, and environmental conditions. The stochastic nature of gene networks has been the focus of many studies (reviewed in [1,2]). There are at least three types of fluctuations that affect gene network dynamics: 1) intrinsic noise or fast fluctuations, 2) extrinsic noise or slow fluctuations, and 3) periodic DNA replicationdependent oscillations [3]. Intrinsic noise arises from the inherent randomness during transcription and translation, key processes for gene expression [4]. Extrinsic noise arises from the factors that universally affect the expression of all genes in a given cell, such as variations in the number of RNA polymerase, ribosome, etc. [1,2]. The third type of fluctuation is due to periodic DNA replication in growing and dividing cells [3].
Nevertheless, gene networks are usually able to perform their regulatory functions under the influence of such disturbances [5], which provokes the question: How do they manage to achieve this remarkable robustness? In control theory, it is known that feedback, a situation in which two (or more) dynamical subsystems are connected in a way that their dynamics are coupled, can make a system resilient towards disturbances [6,7]. A wellknown example of feedback in the context of gene networks is negative autoregulation, in which a transcription factor represses the transcription of its own gene and reduces the effects of noise exerted on the transcription process [810].
Another example is the p53Mdm2 negative feedback, in which p53 transcriptionally activates Mdm2, an E3 ubiquitin ligase, while Mdm2 targets p53 for degradation (Figure 1A) [11]. As one of the most studied tumor suppressor proteins [12,13], p53 plays a key role in repairing DNA damage, arresting cell cycle and, when damage is beyond repair, activating apoptosis (programmed cell death) [14,15]. Therefore, it is important for the cell to regulate p53 robustly, because disturbances may trigger unwanted cell cycle arrest or apoptosis. We have previously shown that the p53Mdm2 negative feedback can reject disturbances and improve robustness under normal (nonstressed) conditions [16]. Intuitively, the negative feedback tries to compensate for changes in the p53 level, so that the impact of any disturbances is offset or at least attenuated.
Figure 1. The p53Mdm2 Feedback Loop. (A) p53 transcriptionally activates Mdm2, while Mdm2 degrades p53. (B) The p53 and Mdm2 levels oscillate upon DNA damage [21]. (C) Under normal conditions, Mdm2 is not suppressed by ATM and the intact feedback rejects disturbances. Upon DNA damage when p53 oscillates, the feedback is suppressed to allow p53 oscillation.
The feedback is suppressed to stop disturbance rejection during DNAdamage. Within minutes of exposure to DNAdamaging agents (UV, Xrays, etc.), the p53 protein level increases rapidly without any significant change in the p53 mRNA level, for p53 is stabilized and no longer degraded by Mdm2 at the normal rate [17,18]. The suppression of the feedback and the stoppage of disturbance rejection allow external factors such as ATM to modulate p53 in response to stress [19,20], which can result in pulses or sustained oscillation of p53 [21] (Figure 1B). Therefore, the feedback mechanism is adaptive – under normal conditions the feedback rejects disturbance to maintain a low steady level of p53 whereas upon DNA damage, the feedback is inactivated to allow pulses or oscillation [16] (Figure 1C).
Interestingly, even though the negative feedback is stronger in normal conditions to reject disturbances, it is well known that feedback can also cause instability, which leads to transient fluctuation (overshoot) and oscillation [7]. Both the strength and the delay of a negative feedback can contribute to instability. As previous measurements showed, the transcriptional regulation of Mdm2 by p53 has a noticeable delay [21] (Figure 1B). Therefore, to maintain a steady p53 level in normal conditions, the p53Mdm2 feedback faces a dilemma: the stronger the feedback is to reject disturbances at the steady state, the more likely the feedback will become unstable and cause transient or sustained fluctuations. The feedback has to carefully balance the tradeoff between steadystate and transient errors for disturbance rejection [22].
Does the p53Mdm2 feedback adopt any strategy to optimize this tradeoff? It is intriguing that in this negative feedback loop Mdm2 downregulates p53 through a posttranslational mechanism (protein degradation), which is not as energy efficient as transcriptional repression because p53 is being produced and actively degraded simultaneously. However, posttranslational regulation has distinct dynamic properties and is relatively faster than transcriptional regulation, so it is conceivable that posttranslational regulations provides an advantage over transcriptional regulation in terms of the robustnessstability tradeoff despite its less energy efficiency. To test this hypothesis, we used techniques from feedback control and adaptive estimation theories to analyze the p53Mdm2 feedback loop.
Results and Discussion
p53Mdm2 feedback model
To study the p53Mdm2 feedback, we started with a previously published p53Mdm2 feedback model that matches experimental measurements [23] and added a term representing ATM (Eqs. 1–3):
where x(t), y(t), and z(t) represent ATM, Mdm2, and p53 levels, respectively. In Eq. 1, since ATM is inactive under normal conditions, the ATM levels are assumed to be low and constant and the rate of change is zero. Eq. 2 is composed of the p53dependent production of Mdm2 (first term), ATMdependent suppression of Mdm2 (second term), and Mdm2 degradation (third term). Eq. 3 is composed of the Mdm2dependent suppression of p53 (first term) and p53 degradation (second term). Note that all the parameter values are supposed to be positive. Following the practice in the published model [23], Eq. 3 does not include the constant basal production rate because it has no effect on the frequency domain and disturbance rejection analysis we will perform next [23]. A discretetime model for the p53Mdm2 feedback system was also built for parameter estimation, which will be described in later sections of the paper.
Feedback reduces steadystate error
Eq. 3 is rather generic and can represent alternative p53 suppression mechanisms by Mdm2. A transcriptional mechanism would only affect p_{yz} in the first term because the strength of transcriptional suppression only depends on the number of Mdm2 molecules (suppressor) but not on the number of p53 protein molecules. In contrast, a posttranslational mechanism would affect both p_{yz} (first term) and p_{z} (second term), because the degradation rate of p53 depends on both the number of p53 molecules and the number of Mdm2 molecules. This generic equation enables us to derive a common set of equations to compare the loop dynamics between transcriptional and translational regulation.
Eqs. 1–3 can be represented as a block diagram using the Laplace transform [24] (Figure 2A). X(s), Y(s), and Z(s) denote the Laplace transforms of x(t), y(t), and z(t) respectively. E(s) is the difference, or error, between the input, p_{xy}X(s), and the output, p_{zy}Z(s). D(s) represents any disturbance exerted on p53. G_{Y}(s) and G_{Z}(s) are the transfer functions that represent the Mdm2 and p53 systems as we showed previously [24]:
Figure 2. Feedback analysis. (A) Block diagram representation of the feedback in Eqs. 1–3. X(s), Y(s) and Z(s) are the Laplace transforms of x(t), y(t), and z(t). G_{Y}(s) and G_{Z}(s) are the transfer functions. E(s) is the Laplace transform of the error e(t) between the input and output. D(s) represents the disturbance. (B) Time responses (unit step response) for a secondorder system with different damping ratios (p_{z} = p_{y} = 0.01 min^{1}). (C) As the damping ratio (ζ) increases from 0 to near 1 the %OS decreases from 100% to near 0%. (D) Block diagram representation of the discretetime feedback model. X(z), Y(z), and Z(z) are the ztransforms of x(i), y(i), and z(i). G_{Y}(z) and G_{Z}(z) are the transfer functions. E(z) is the transform of the error e(i) between the input and output. D(z) represents the disturbance.
From Figure 2A, E(s) can be expressed as [16]:
From the same figure, Z(s) can be expressed as:
Substituting Eq. 6 for Z(s) into Eq. 5, we obtain:
The second term in Eq. 7 represents the contribution to E(s) from D(s), the Laplace transform of the disturbance signal. We denote this term as E_{D}(s) and its corresponding time domain function as e_{D}(t):
Using the final value theorem and assuming a step disturbance (D(s) = 1/s), we can determine the steadystate error due to the disturbance as follows:
According to Eq. 9, increasing either p_{yz} or p_{z} will reduce the steadystate error, because both terms are only in the denominator. Therefore, either a stronger negative feedback through the Mdm2 suppression of p53 (p_{yz}) or less stable p53 (p_{z}) can reduce the steadystate error. Interestingly, as stated previously, posttranslational degradation of p53 by Mdm2 increases both p_{yz} and p_{z} (degradation rate), while a hypothetical transcriptional suppression would only increase p_{yz}. Hence posttranslational suppression of p53 may be more efficient at reducing steadystate error than transcriptional suppression, even though it is less energy efficient.
Tradeoff between steadystate and transient errors
The feedback has to minimize its transient response to disturbances in addition to steadystate error; otherwise, a temporary overshoot of the p53 level may trigger unintended effects. We therefore examine the percentage overshoot (%OS), which is the amount that the p53 level transiently overshoots the final steadystate level (expressed as a percentage of the final value). Because %OS is a function of the damping ratio (ζ), we derive a secondorder transfer function representative of the block diagram in Figure 2A to determine its damping ratio (ζ). Substituting Eq. 5 for E(s) in Eq. 6, the transfer function G(s), which directly relates the input X(s) to the output Z(s), can be expressed as:
From Eq. 10, the natural frequency and damping ratio can be expressed as [24]:
And the %OS is given by [7]:
The typical time responses (unit step responses) for a secondorder system with different damping ratios are shown in Figure 2B (p_{z} = p_{y} = 0.01 min^{1}). The %OS, or overshoot of the p53 level, decreases when the damping ratio increases (Figure 2C).
According to Eq. 12, increasing p_{yz} monotonically reduces the damping ratio and increases %OS, for p_{yz} only appears in the denominator. Since, transcriptional suppression can only influence p_{yz}, a transcriptional negative feedback is limited by the tradeoff between steadystate error and transient overshoot – increasing the strength of the negative feedback reduces steadystate error at the expense of increasing transient overshoot.
On the other hand, how does p_{z} affect the overshoot? If we take the partial derivative of the damping ratio in regard to p_{z}, we get:
The secondorder partial derivative of Eq. 14 shows that ζ reaches the minimum value at . When , ζ decreases with p_{z}; when , ζ increases with p_{z}. This suggests that when , increasing p_{z} can decrease both the steadystate error and overshoot (which is inverse to the damping ratio). This insight potentially explains why cells choose to spend extra energy producing and then actively degrading p53 under normal conditions – it rejects disturbances by reducing both the steadystate error and the transient overshoot. Furthermore, posttranslational suppression influences both p_{yz} and p_{z}, because the Mdm2mediated degradation rate of p53 depends on both Mdm2 and p53 concentrations. This suggests that posttranslational suppression can achieve a better tradeoff than transcriptional suppression, which can only influence p_{yz}.
To quantitatively verify these analytical insights, we calculated the steadystate error, damping ratio, and %OS using published, experimentally measured parameter values (p_{yz} = 0.8 h^{1}[23], p_{zy} = 0.8 h^{1}[23], p_{y} = 2.0 h^{1} (halflife ≈ 20 min) [25], p_{z} = 2.0 h^{1} (halflife ≈ 20 min) [26]). As p_{yz} increases from 0.5 to 3.0, the steadystate error decreases while the %OS increases (the damping ratio decreases) (Figure 3A, left panel). Hence there is a tradeoff between steadystate error and transient overshoot (Figure 3A, right panel), consistent with the previous analysis. Any further decrease of the steadystate error from the system operating point (p_{yz} = 0.8) will have to pay a hefty penalty for transient overshoot. On the other hand, as p_{z} is increased from 0.5 to 3.0, the %OS initially increases but then decreases (Figure 3B, left panel), exactly as Eqs. 14 and 15 have indicated. The system operates at a point (p_{y} = 2.0, p_{z} = 2.0) where , which allows the system to reduce both the steadystate error and %OS through p_{z} (Figure 3B, right panel). The combined effect of p_{yz} and p_{z} is shown in Figure 3C and D. The steadystate error can be decreased by either increasing p_{yz} or p_{z} (Figure 3C), but increasing p_{yz} can increase %OS (Figure 3D), as shown by the arrows in the figures. Therefore, by modulating both p_{yz} and p_{z}, posttranslational suppression can achieve better steadystate error and %OS than transcriptional suppression, which only modulates p_{yz}.
Figure 3. Tradeoff between steadystate error and transient overshoot. (A) Steadystate error and %OS (left panel) and the tradeoff curve (right panel) when p_{yz} is increased from 0.5 to 3.0. The orange cross denotes the operating point (p_{yz} = 0.8). (B) Steadystate error and %OS (left panel) and the tradeoff curve (right panel) when p_{z} is increased from 0.5 to 3.0. The orange cross denotes the operating point (p_{z} = 2.0). (C) Increasing p_{yz} or p_{z} can reduce the steadystate error. (D) Increasing p_{yz} can increase %OS. A better operating point can be achieved by simultaneously increasing p_{z}.
Interestingly, the fact that p_{z} can improve both steadystate error and %OS seems to suggest that cells should keep increasing p_{z} to achieve ever better robustness. However, it is worth noting that there is another tradeoff factor that ultimately comes into the picture, which is the energy cost. Higher p_{z} means that the cells have to spend more resources to rapidly produce and then degrade p53 in a “futile” cycle, so the degradation rate cannot be increased indefinitely. The cell has to balance between robustness and energy spending to achieve the optimal operating point.
Altogether, our analytical and quantitative analyses suggest that cells spend extra energy to produce and degrade p53 simultaneously under normal conditions in order to maintain a more robust p53 level. By modulating both p_{yz} and p_{z}, posttranslational suppression of p53 enables the p53Mdm2 feedback to achieve a better tradeoff by reducing both steadystate errors and transient overshoots. On the contrary, transcriptional suppression only modulates p_{yz}, which does not improve the overall tradeoff.
Robustness analyses with pole plots
As mentioned at the beginning of the paper, gene networks fluctuate over time under the influence of extrinsic and intrinsic noise. The range of p_{yz} and p_{z} in Figure 3 demonstrated the robustness tradeoff between steadystate and transient disturbance rejection, but it will be useful to examine the tradeoff within the range of parameter fluctuation that a realistic cell has to experience. To estimate the timevarying parameter values from timeseries data of the p53Mdm2 network [21], we constructed a corresponding discretetime model, which can be described by Eqs. 16–18:
where x(i), y(i), and z(i) represent ATM, Mdm2, and p53 levels respectively. As before, the ATM level is assumed to be low and constant under normal conditions. w_{xy} represents the suppression of Mdm2 by ATM, w_{yz} represents the suppression of p53 by Mdm2, w_{zy} represents the transcriptional activation of Mdm2 by p53, w_{y} represents the stability of Mdm2 (1w_{y} represents Mdm2 degradation), and w_{z} represents the stability of p53 (1w_{z} represents p53 degradation). Note that w_{y} and w_{z} represent stability rather than degradation, opposite to p_{y} and p_{z}. Block diagram representation of the discretetime feedback model is shown in Figure 2D. Consistent with the continuous model, the discretetime model confirms that the steadystate error can be decreased by increasing w_{yz} or decreasing w_{z} (Eq. 19, also see Eq. S11 in Additional file 1). Note that decreasing w_{z} means increasing p53 degradation and corresponds to increasing p_{yz}.
Additional file 1. Supplementary document.
Format: DOC Size: 252KB Download file
This file can be viewed with: Microsoft Word Viewer
We used the timeseries experimental data [21] and an NLMS adaptive filter [27,28] to track the timevarying parameters of the p53Mdm2 model (see Additional file 1) (Figure 4A). The published experimental data [21] and MATLAB code used to estimate the parameter ranges are included in Additional files 2, 3, 4 and 5. Based on the estimates, we examined how w_{yz} and w_{z} affect %OS.
Figure 4. Stability analysis. (A) Mean and standard deviation of each parameter identified by the adaptive filter algorithm: w_{y}: 0.8737±0.0830, w_{z}: 0.3627±0.0825, w_{zy}: 0.2662±0.0343, w_{yz}: 0.1656±0.1169. (B) Pole map under the DNA damage condition when the feedback is weakened. The poles are close to the real axis and the corresponding damping ratios range mostly from 0.8 to 1.0. (C) Poles are located higher with bigger imaginary components and lower damping ratios (0.4 – 0.8) when w_{yz} is increased. (D) Poles are closer to the real axis with bigger damping ratios compared to (C) when w_{z} is decreased proportionally. T in the figure denotes the sampling period (0.11 hr). The numbers next to the arrow indicate the corresponding damping ratios.
Format: TXT Size: 5KB Download file
Format: TXT Size: 5KB Download file
Additional file 4. p53Mdm2 parameter tracking using NLMS (Mdm2 estimation).
Format: DOC Size: 26KB Download file
This file can be viewed with: Microsoft Word Viewer
Additional file 5. p53Mdm2 parameter tracking using NLMS (p53 estimation).
Format: DOC Size: 26KB Download file
This file can be viewed with: Microsoft Word Viewer
Using Ztransform [29], we first derived a transfer function, G(z), which represents the feedback system shown as the block diagram in Figure 2D (see Additional file 1):
To investigate the transient behavior and stability of G(z), we plotted its poles using the MATLAB Robust Control Toolbox (Figure 4BD), which can be reproduced using the supplementary MATLAB file (Additional file 6). For each map, 100 points were calculated based on Monte Carlo sampling of the estimated parameter range in Figure 4A. We first plotted the poles for the DNA damage condition, under which the Mdm2 suppression of p53 (the negative feedback) is weakened to stop disturbance rejection (Figure 4B). Consistent with the weakening of the feedback, the plot shows that the poles are mostly real and the system is stable, with damping ratios ranging from 0.8 to 1.0.
Additional file 6. p53Mdm2 discrete uncertain model: stability analysis.
Format: DOC Size: 29KB Download file
This file can be viewed with: Microsoft Word Viewer
Western blot measurements of the total and ubiquitinated p53 levels showed that the suppression of p53 is 3.67 fold higher under the normal condition compared to the DNA damage condition [30]. We first increased w_{yz} by 3.67 fold (from 0.1656±0.1169 to 0.6708±0.1169) to evaluate how transcriptional suppression will affect transient overshoot and stability. The increased w_{yz} values shifted poles higher on the map compared to the DNA damage condition (Figure 4C). The bigger imaginary components of the poles indicate that the damping ratios are lower (0.4 – 0.8) and the corresponding %OS values are higher for the system, therefore confirming our previous conclusion that transcriptional suppression reduces the steadystate error at the expense of transient overshoot and stability.
However, if the increase of w_{yz} is accompanied by a decrease of w_{z} (stability of p53) by 3.67 fold (0.3627±0.0825 to 0.0998±0.0825), which approximates posttranslational suppression, the poles are located closer to the real axis. The smaller imaginary components of the poles indicate that the damping ratios are greater (the %OS values are smaller). Therefore, posttranslational suppression can reduce the steadystate error without a hefty penalty of transient overshoot (%OS) and stability, unlike transcriptional suppression.
Posttranslational regulation enables faster responses with predictable level shifts
Posttranslational regulation has another benefit over transcriptional regulation – it enables faster responses with more consistent amplitude. This is biologically significant, as cells must react quickly to external modulators (e.g. ATM) and stress conditions with a predictable shift of p53 levels. From Eq. 3, the step response of the p53Mdm2 feedback can be shown as
with the assumption that y(t) has a constant value of Y. Eq. 21 suggests that p_{z} but not p_{yz} determines the response time (the time needed for z(t) to reach the half steadystate value) (Figure 5A and B). However, the increase in speed comes at a cost; it decreases the response amplitude as shown in Figure 5B. Interestingly, increasing both p_{yz} and p_{z} allows the system to achieve a faster response time with constant steadystate amplitude (Figure 5C), because the respective increases of p_{yz} and p_{z} offset each other at the steady level (Eq. 21). Therefore, by modulating both p_{yz} and p_{z}, posttranslational suppression can generate more prompt and consistent p53 response to stress conditions. Indeed, cells respond to DNA damage by modulating the posttranslational degradation of p53.
Figure 5. Step responses of p53. (A)p_{yz} does not influence the response time. (B)p_{z} decreases the response time but also the amplitude of the response. (C) Increasing p_{yz} while decreasing p_{z} enables faster response time with constant amplitude.
Validation by a nonlinear, mechanistic model
So far, our analyses have been performed using the linear p53MDM2 feedback model (Eq. 1–3) modified from GevaZatorsky et al. [23]. Even though this model matches experimental measurements [23], it does not capture the nonlinear aspects of the negative feedback loop. Therefore, the linear model, and hereby its analyses, can only be applied to the first order approximation.
To test whether our findings are still valid when nonlinear effects are taken into consideration, we built a mechanistic model of the p53MDM2 feedback [Eqs. 22–25]:
where, [mdm2_{mRNA}], [Mdm2], [p53_{mRNA}], and [P53] denote the mRNA and protein levels of p53 and Mdm2. β_{1} and β_{3} denotes transcription rates, and β_{2} and β_{4} denote translation rates of Mdm2 and p53. α_{s} (s = 1,2,3,4) denote mRNA and protein degradation rates. Transcriptional activation of Mdm2 by p53 is modeled by a Hill function, where n is the hill coefficient and K_{1} is the dissociation constant. Mdm2mediated p53 ubiquitination and degradation are modeled by a MichaelisMenten (MM) function as described by Xu et al. [31]. γ is the reaction rate of p53 ubiquitination by Mdm2, and K_{2} is the saturation constant.
To measure disturbance rejection, p53 production was increased by 1% from its steadystate level (simulated by a step function). Steadystate error and %OS were then calculated based on the following equations:
where OS is the overshoot level, SS_{old} is the p53 steadystate level before perturbation, and SS_{new} is the p53 steadystate level after perturbation (Figure 6A).
Figure 6. Disturbance rejection simulation of a mechanistic model. (A) Simulation of a step function disturbance to p53 production. The equations illustrate how the steadystate error and %OS are calculated. (B) Steadystate error and %OS (top panel) and the tradeoff curve (bottom panel) when γ increases. (C) Plots of %OS and steadystate error vs. γ with varying Mdm2 levels. (D) Plots of %OS vs. γ with varying time delay. (E) Time delay can cause sustained oscillation (β_{1}=1, β_{2}=1, β_{3}=1, β_{4}=1, α_{1} =1, α_{2} =0.01, α_{3} =1, α_{4} =0.01, n =2, K_{1} =10, K_{2} =100, γ=10^{2} 10^{2}).
When γ, the degradation rate of p53 by Mdm2mediated ubiquitination, increases, the %OS initially increases but then decreases, which allows the system to reduce both the steadystate error and %OS through gamma (Figure 6B). This plot is similar to the previous analysis with the linear model when p_{z} is increased (Figure 3B). Therefore, the mechanistic model confirmed the insight from the linear model that posttranslational suppression enables the p53Mdm2 negative feedback to reduce both steadystate error and %OS.
We explored the parameter space to examine the different operating regions of the nonlinear model, and found that the above tradeoff trend is generally preserved even though the absolute values of steadystate error and %OS vary (Figure 6C). This is probably due to the fact that the model is roughly piecewise linear at each operating point when given a modest disturbance, so that higherorder effects do not dominate the firstorder behavior predicted by the linear model.
Besides nonlinearity, the linear model is also overly simplistic in terms of another critical aspect of the p53Mdm2negative feedback, which is time delay. Time delay in a negative feedback loop can decrease stability, increase %OS and cause sustained oscillation. To examine how time delay affects disturbance rejection, we added explicit delay terms between transcription and translation [Eqs. 28–31]:
where τ denotes the time delay between transcription and translation.
Simulations with varying time delay reveal that delay does not affect steadystate error but increases %OS, which is consistent with the proposition that time delay decreases stability of negative feedback loops (Figure 6D). A longer delay also increases the likelihood of sustained oscillation (Figure 6E).
Stochastic simulation
So far, the deterministic p53Mdm2 models have shown how the negative feedback helps reject external disturbances. However, how does the feedback affect intrinsic, stochastic noise? To understand this issue, we constructed a mechanistic, kinetic model and performed stochastic simulations using the MATLAB SimBiology Toolbox (see Additional file 1) (Figure 7A). The mean and standard deviation of the steadystate p53 levels are calculated based on stochastic runs (Figure 7B). We varied γ, the degradation rate of Mdm2ubiquitinated p53, while keeping the p53 level constant, to investigate how the negative feedback affects p53 variance (Figure 7C). The stochastic noise was then calculated as the variance normalized by the mean (Figure 7D). These stochastic simulations reveal that stochastic noise on p53 increases when γ increases. This suggests that while Mdm2mediated p53 degradation improves disturbance rejection, it pays the price of amplifying stochastic noise. Hence, the p53Mdm2 negative feedback loop has to balance the tradeoff between disturbance rejection and stochastic noise.
Figure 7. Stochastic simulation of the feedback loop. (A) p53 levels during repeated stochastic runs. (B) Mean and standard deviation of the steadystate p53 levels calculated from stochastic runs. (C) Mean and standard deviation of p53 with varying γ. We maintained a relatively constant p53 level by adjusting its natural degradation rate accordingly. (D) Variance normalized by mean increases with γ.
Conclusions
In summary, our analyses indicate that cells maintain robust p53 levels and reject disturbances by simultaneously producing and degrading p53, even though this process is more energy intensive. Within the p53Mdm2 negative feedback loop, posttranslational suppression of p53 by Mdm2 achieves a better tradeoff between steadystate and transient errors than transcriptional suppression, which potentially explains why the former has been experimentally observed in the cell. Furthermore, posttranslational suppression enables p53 to respond faster to stress conditions with a more predictable level shift. Understanding these nuances allows us to appreciate the complexity of regulatory networks, which will potentially lead to better therapeutics.
Besides ubiquitination, p53 activity is also heavily regulated by phosphorylation and nucleocytoplasmic shuttling [3235]. Interestingly, these posttranslational processes provide similar beneficial tradeoff between steadystate errors and overshoot, while being faster and less energy intensive than protein degradation. It is probably not a coincidence then that p53 has multiple phosphorylation sites and is shuttled alongside Mdm2 between nucleus and cytoplasm. Therefore, protein degradation is only one of several posttranslational mechanisms that enhance the robustness of the system.
Increasing degradation rather than reducing production might be a common strategy evolved by biological systems for robustness. For example, hematopoietic stem cells continuously go through apoptosis [36], which seems energyinefficient and futile as the protein degradation we discussed. However, our analysis would suggest that regulating apoptosis rather than cell division may enable the stem cell population to become more robust to disturbances and respond faster to changes. The fact that biological systems employ mechanisms for robustness at many different levels raises an interesting question  how does robustness at each level contribute to the overall robustness of the whole system? Undoubtedly challenging, attempts to answer this question will help unravel the underlying design principles of complex biological systems.
Methods
Computational methods
Ordinary differential and difference equations were used for physicsbased modeling of the p53Mdm2 feedback loop (see Additional file 1 for the derivation of the p53Mdm2 discretetime model from underlying physics). Parameter ranges of the discretetime model were estimated using the Normalized Least Mean Squares (NLMS) method detailed in Additional file 1. The experimental data [21] and MATLAB (Mathworks, USA) codes are provided as (Additional files 1, 2, 3, 4 and 5). Steadystate and transient error analysis was performed using Laplace and Zdomain analysis (see Additional file 1 for steadystate error analysis using the discretetime model). The Monte Carlo method and MATLAB Robust Control Toolbox (Mathworks, USA) were used for the estimationbased robustness analysis (see Additional file 6).
Image extraction and fluorescence quantification
285 Image frames were extracted from the video file [21] and the fluorescence quantification of p53 and Mdm2 was carried out using the LabVIEW Vision Assistant 2010 (National Instruments, USA). We manually marked the location of each cell nucleus in each frame and 285 data points were obtained for each protein.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
YS conceived of and designed the study, performed linear, discretetime and robustness analyses and drafted the manuscript. KC carried out nonlinear and stochastic analyses. AHS assisted in the adaptive filter analysis. BH participated in designing the robustness analysis. XS conceived of the study, participated in its design, and helped to draft the manuscript. All authors have read and approved the manuscript.
Acknowledgements
This work was supported by National Institutes of Health grant R01GM95990 and National Science Foundation grants NSF 1137269, CCF1011918 and CCF0942936.
References

Kaern M, Elston TC, Blake WJ, Collins JJ: Stochasticity in gene expression: From theories to phenotypes.
Nat Rev Genet 2005, 6:451464. PubMed Abstract  Publisher Full Text

Raj A, van Oudenaarden A: Nature, nurture, or chance: Stochastic gene expression and its consequences.
Cell 2008, 135:216226. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Rosenfeld N, Young JW, Alon U, Swain PS, Elowitz MB: Gene regulation at the singlecell level.
Science 2005, 307:19621965. PubMed Abstract  Publisher Full Text

Dunlop MJ, Cox RS, Levine JH, Murray RM, Elowitz MB: Regulatory activity revealed by dynamic correlations in gene expression noise.
Nat Genet 2008, 40:14931498. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Jen E: Robust design : Repertoire of biological, ecological, and engineering case studies. New York: Oxford University Press; 2005.

Cosentino C, Bates D: Feedback control in systems biology. Boca Raton, Florida: CRC Press; 2012.

Åström KJ, Murray RM: Feedback systems: An introduction for scientists and engineers. Princeton: Princeton University Press; 2008.

Becskei A, Serrano L: Engineering stability in gene networks by autoregulation.
Nature 2000, 405:590593. PubMed Abstract  Publisher Full Text

Simpson ML, Cox CD, Sayler GS: Frequency domain analysis of noise in autoregulated gene circuits.
Proc Natl Acad Sci U S A 2003, 100:45514556. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Dublanche Y, Michalodimitrakis K, Kummerer N, Foglierini M, Serrano L: Noise in transcription negative feedback loops: Simulation and experimental analysis.
Mol Syst Biol 2006, 2:41. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Piette J, Neel H, Marechal V: Mdm2: Keeping p53 under control.
Oncogene 1997, 15:10011010. PubMed Abstract  Publisher Full Text

Vogelstein B, Lane D, Levine AJ: Surfing the p53 network.
Nature 2000, 408:307310. PubMed Abstract  Publisher Full Text

Levine AJ, Oren M: The first 30 years of p53: Growing ever more complex.
Nat Rev Cancer 2009, 9:749758. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Kaina B: DNA damagetriggered apoptosis: Critical role of DNA repair, doublestrand breaks, cell proliferation and signaling.
Biochem Pharmacol 2003, 66:15471554. PubMed Abstract  Publisher Full Text

Roos WP, Kaina B: DNA damageinduced cell death by apoptosis.
Trends Mol Med 2006, 12:440450. PubMed Abstract  Publisher Full Text

Shin YJ, Hencey B, Lipkin SM, Shen X: Frequency domain analysis reveals external periodic fluctuations can generate sustained p53 oscillation.
PLoS One 2011, 6:e22852. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Maltzman W, Czyzyk L: UV irradiation stimulates levels of p53 cellular tumor antigen in nontransformed mouse cells.
Mol Cell Biol 1984, 4:168994. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Price BD, Calderwood SK: Increased sequencespecific p53DNA binding activity after DNA damage is attenuated by phorbol esters. BASINGSTOKE: ONCOGENE; 1993.
8: 3055

Kastan MB, Zhan Q, ElDeiry WS, Carrier F, Jacks T, et al.: A mammalian cell cycle checkpoint pathway utilizing p53 and GADD45 is defective in ataxiatelangiectasia.
Cell 1992, 71:587597. PubMed Abstract  Publisher Full Text

Gannon HS, Woda BA, Jones SN: ATM phosphorylation of Mdm2 Ser394 regulates the amplitude and duration of the DNA damage response in mice.
Cancer Cell 2012, 21:668679. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

GevaZatorsky N, Rosenfeld N, Itzkovitz S, Milo R, Sigal A, et al.: Oscillations and variability in the p53 system.
Mol Syst Biol 2006, 2:2006.0033. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Chandra FA, Buzi G, Doyle JC: Glycolytic oscillations and limits on robust efficiency.

GevaZatorsky N, Dekel E, Batchelor E, Lahav G, Alon U: Fourier analysis and systems identification of the p53 feedback loop.
Proc Natl Acad Sci U S A 2010, 107:1355013555. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Shin YJ, Bleris L: Linear control theory for gene network modeling.
PLoS One 2010, 5:e12785. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Stommel JM, Wahl GM: Accelerated MDM2 autodegradation induced by DNAdamage kinases is required for p53 activation.
EMBO J 2004, 23:154756. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Ju J, Schmitz JC, Song B, Kudo K, Chu E: Regulation of p53 expression in response to 5fluorouracil in human cancer RKO cells.
Clin Cancer Res 2007, 13:42454251. PubMed Abstract  Publisher Full Text

Shin YJ, Sayed AH, Shen X: Adaptive models for gene networks.
PloS one 2012, 7:e31657. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Franklin GF, Powell JD: Digital control of dynamic systems. Reading, Mass.: AddisonWesley Pub. Co.; 1980.

Cheng Q, Cross B, Li B, Chen L, Li Z, et al.: Regulation of MDM2 E3 ligase activity by phosphorylation after DNA damage.
Mol Cell Biol 2011, 31:49514963. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Xu L, Qu Z: Roles of protein ubiquitination and degradation kinetics in biological oscillations.
PloS one 2012, 7(4):e34616. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

O'Keefe K, Li H, Zhang Y: Nucleocytoplasmic shuttling of p53 is essential for MDM2mediated cytoplasmic degradation but not ubiquitination.
Mol Cell Biol 2003, 23:6396405. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Tao W, Levine AJ: Nucleocytoplasmic shuttling of oncoprotein Hdm2 is required for Hdm2mediated degradation of p53.
Proc Natl Acad Sci 1999, 96:30773080. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Ashcroft M, Kubbutat MH, Vousden KH: Regulation of p53 function and stability by phosphorylation.
Mol Cell Biol 1999, 19:17518. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Kazuyasu S, Herrera JE, Shin’ichi S, Toru M, Michael B, Alex V, Anderson CW, Anderson CW, Ettore A: DNA damage activates p53 through a phosphorylation–acetylation cascade.
Genes Dev 1998, 12(18):28312841. PubMed Abstract  Publisher Full Text  PubMed Central Full Text

Alenzi FQ, Ahmad SY, Alenazi BQ, Salem ML, AlJabri AA, et al.: The haemopoietic stem cell: Between apoptosis and self renewal.
Yale J Biol Med 2009, 82:718. PubMed Abstract  PubMed Central Full Text