• We are sorry, but NCBI web applications do not support your browser and may not function properly. More information
Logo of ploscompComputational BiologyView this ArticleSubmit to PLoSGet E-mail AlertsContact UsPublic Library of Science (PLoS)
PLoS Comput Biol. Mar 2011; 7(3): e1002010.
Published online Mar 10, 2011. doi:  10.1371/journal.pcbi.1002010
PMCID: PMC3053318

Noise Contributions in an Inducible Genetic Switch: A Whole-Cell Simulation Study

Peter Csermely, Editor

Abstract

Stochastic expression of genes produces heterogeneity in clonal populations of bacteria under identical conditions. We analyze and compare the behavior of the inducible lac genetic switch using well-stirred and spatially resolved simulations for Escherichia coli cells modeled under fast and slow-growth conditions. Our new kinetic model describing the switching of the lac operon from one phenotype to the other incorporates parameters obtained from recently published in vivo single-molecule fluorescence experiments along with in vitro rate constants. For the well-stirred system, investigation of the intrinsic noise in the circuit as a function of the inducer concentration and in the presence/absence of the feedback mechanism reveals that the noise peaks near the switching threshold. Applying maximum likelihood estimation, we show that the analytic two-state model of gene expression can be used to extract stochastic rates from the simulation data. The simulations also provide mRNA–protein probability landscapes, which demonstrate that switching is the result of crossing both mRNA and protein thresholds. Using cryoelectron tomography of an E. coli cell and data from proteomics studies, we construct spatial in vivo models of cells and quantify the noise contributions and effects on repressor rebinding due to cell structure and crowding in the cytoplasm. Compared to systems without spatial heterogeneity, the model for the fast-growth cells predicts a slight decrease in the overall noise and an increase in the repressors rebinding rate due to anomalous subdiffusion. The tomograms for E. coli grown under slow-growth conditions identify the positions of the ribosomes and the condensed nucleoid. The smaller slow-growth cells have increased mRNA localization and a larger internal inducer concentration, leading to a significant decrease in the lifetime of the repressor–operator complex and an increase in the frequency of transcriptional bursts.

Author Summary

Expressing genes in a bacterial cell is noisy and random. A colony of bacteria grown from a single cell can show remarkable differences in the copy number per cell of a given protein after only a few generations. In this work we use computer simulations to study the variation in how individual cells in a population express a set of genes in response to an environmental signal. The modeled system is the lac genetic switch that Escherichia coli uses to find, collect, and process lactose sugar from the environment. The noise inherent in the genetic circuit controlling the cell's response determines how similar the cells are to each other and we study how the different components of the circuit affect this noise. Furthermore, an estimated 30–50% of the cell volume is taken up by a wide variety of large biomolecules. To study the response of the circuit caused by crowding, we simulate the circuit inside a three-dimensional model of an E. coli cell built using data from cryoelectron tomography reconstructions of a single cell and proteomics data. Correctly including random effects of molecular crowding will be critical to developing fully dynamic models of living cells.

Introduction

Transcriptional and translational regulatory networks control the phenotype of modern cells, regulating gene expression in response to changing environmental conditions and/or biological stimuli. It has been well established that intrinsic noise in gene regulation results from the discrete biochemical nature of the process [1]. There is also an extrinsic component to the total noise arising from cell-to-cell variation in the number of copies of the transcription and translation machinery (transcription factors, RNA polymerases, ribosomes, etc) [2][4]. Stochastic noise can lead to different phenotypic outcomes for a cellular population and, in certain fluctuating environments, the resulting heterogeneous population can be more optimal for growth than would be a population containing a single phenotype [5], [6].

Theoretical modeling of stochasticity in gene expression has been a topic of intense study in the last decade and has greatly increased our understanding of the effect that statistical noise has on gene regulation (for reviews see [7][11]). Without detailed information regarding spatial heterogeneity within a cell, models of stochastic gene expression are typically expressed in terms of the chemical master equation (CME), which describes the time evolution of the probability for a chemical system to be in a given state [12]. Various analytical methods including moment generating functions [1], [3], [13], the Langevin and Fokker-Planck equations [14], linear noise approximation [4], and many-body theory [15] are used to study such models of gene expression. Computer simulations, usually based on a variant of Gillespie's stochastic simulation algorithm (SSA) [16] are also widely employed to analyze gene network models that are too complex to be amenable to analytical modeling [17], [18].

Such theoretical studies have predicted and experimental measurements have shown [2], [19][23] that populations of cells can be quite heterogeneous, even when starting from an initially identical state. The large variance in the population distribution is usually ascribed to bursting in the process of gene transcription. Two models have been developed which can be used as a framework for quantitatively analyzing population distributions to infer the underlying gene expression kinetics.

The burst model (Figure 1A) of Friedman et al. [24] is based on the assumption that an mRNA's lifetime is short compared with that of its protein product. In that case, proteins will be produced in independent bursts with exponentially distributed sizes. The solution to the stationary probability distribution of protein An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e001.jpg in the continuous CME formulation of the model was shown to be the Gamma distribution An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e002.jpg where the An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e003.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e004.jpg parameters were interpreted to be the frequency of transcriptional bursts relative to the protein lifetime and the mean number of proteins produced per burst, respectively.

Figure 1
Three models for stochastic gene expression.

Shahrezaei and Swain [25] further developed the analytical theory of gene expression, by deriving not only the time-dependent probability distribution for the burst model, but also the steady-state distribution for a two-state model of gene expression (Figure 1B; three-stage model in their nomenclature). In the two-state model a gene alternates between transcriptionally active and inactive states with constant rates. Their analytical distributions show that in addition to large variance within a population, bimodality can appear when transitions between the active and inactive states are slow. A similar model has also been used to analyze the switching behavior of a population due to rare large events versus the cumulative effect of many small events [26].

Computational modeling can greatly assist in understanding genetic systems where complexity exceeds the capacity of analytical solutions. In a model for an inducible genetic switch incorporating more of the complexity present in real biological systems (Figure 1C), the transitions between the active and inactive transcriptional states are no longer constant but depend upon an external inducer likely in a nonlinear manner. The positive feedback (PFB) loop changes the network topology by introducing an additional regulatory link. Both of these differences provide additional sources of noise in the circuit that may affect the probability distributions. Combining computer modeling of a complete genetic circuit with analysis using simplified analytical models can help to provide an overall picture of the dynamics of such a system.

Further complexity in modeling real biological systems comes from the spatial heterogeneity within a cell and molecular crowding in the in vivo environment. It is becoming apparent that the cell is not a well-stirred system [27][29]. Studies using cryoelectron tomography techniques [30][34] have revealed that individual macromolecules are not necessarily uniformly distributed inside the cell, but may be clustered in a spatially dependent manner. Spatial organization can affect reaction kinetics by increasing local concentrations of reactants and enzymes. Additionally, crowding and non-specific molecular interactions in the in vivo environment can lead to anomalous subdiffusive behavior for macromolecules, as measured experimentally [35], [36] and by computational modeling of bacterial cytoplasmic environments [37][39]. Accounting for spatial heterogeneity is a challenge to computational biology that must eventually be met and several such modeling studies have been undertaken [37][44].

Stochastic modeling of gene expression circuits in a three-dimensional bacterial cell poses several difficulties, both computational and informational in nature. Recently a “lattice microbe” method [37] was developed using GPU (graphics processing unit) computational accelerators to simulate diffusion of macromolecules within a modeled Escherichia coli cell packed with a distribution of obstacles according to reported proteomics data. It implemented a multiparticle reaction-diffusion algorithm on a three-dimensional lattice to perform simulations of cell-scale systems. With the lattice microbe method one can observe anomalous diffusion of macromolecules and track diffusive-reactive processes over the timescale of the cell cycle, with spatial resolution from 2–16 nm. On the informational side, painstaking efforts must be undertaken to obtain parameters for the models. Kinetic parameters, which are often obtained under in vitro conditions, must be validated by comparing modeling results to published experiments. Recent time-lapse fluorescence microscopy experiments have been able to track dynamic behavior for individual macromolecules in vivo [21], [45], providing an additional source for model parameters. Parameters obtained from in vivo single-molecule experiments are uniquely suited for stochastic modeling, as they provide population distributions not simply mean values from ensemble measurements. Equally importantly such parameters are measured under in vivo conditions and incorporate the effects of the cellular environment. Also, super-resolution imaging studies [46][49] provide further spatial information to complement the cryoelectron tomography data.

We present here a computational study of gene expression noise in the inducible genetic switch shown in Figure 1C using both well-stirred and spatially resolved models. Spatial models of E. coli cells were constructed to approximate cytoplasmic crowding under both rapid and slow growth phenotypes, with the latter being based on data from cryoelectron tomography [50]. Both spatial models were simulated using the lattice microbe method [37]. The genetic switch was based on the well-characterized E. coli lactose utilization system, parameterized using measurements from a recent series of in vivo single-molecule fluorescence studies [21], [22], [51] as well as published in vitro rate constants. We report the contributions to intrinsic noise from the regulatory elements of the inducible genetic circuit as well as the extrinsic noise due to in vivo crowding. Using the slow-growth model we investigate the effect of using experimentally determined cellular architecture in reaction-diffusion models, with implications for effects due to cell growth. Comparing the noise from the inducible genetic switch to the bursting and two-state models described above (Figure 1A,B), we consider what improvements in both modeling and experimental efforts are needed to develop stochastic models of gene expression with predictive power regarding phenotype switching and heterogeneity in cellular populations.

Methods

Lac circuit kinetic model

Lactose uptake in Escherichia coli

The lactose utilization system in E. coli is a model system for studying inducible genetic circuits [52][58]. The overall genetic system is described in Figure 2. Briefly, the lac repressor (LacI; An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e005.jpg in model annotation) [59][61] binds to the lac operator (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e006.jpg in model annotation) upstream of the DNA encoding for the genes responsible for lactose uptake and metabolism, repressing their expression in the absence of lactose. In the presence of lactose or another inducer (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e007.jpg in model annotation), LacI binds the inducer preferentially and is prevented from binding to the operator region allowing expression of the proteins in the lac operon. One protein in the operon, lactose permease (LacY; An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e008.jpg in model annotation), establishes positive feedback in the circuit by inserting into the membrane and actively transporting lactose into the cell, ensuring that LacI remains sequestered; the cell switches to the induced state. Theoretical and experimental studies have investigated the behavior of lac system and shown it to be stochastic, depending on random fluctuations to switch between the off and on states [62], [63]. We assumed the same overall kinetic structure for our model of the lac system as Stamatakis and Mantzaris [63], but where possible derived the rate parameters from single molecule in vivo experiments. The reactions and stochastic rate constants, derived using a well-stirred approximation, are given in Table 1.

Figure 2
Overview of the lac genetic circuit in E. coli.
Table 1
Reactions and rate constants used in the stochastic model of the lac circuit.

Lac operon regulation: Activation and inactivation of transcription

The regulatory behavior of the lac circuit results from the binding of the repressor to the lac operator, thereby inhibiting transcription initiation. There are three possible inducer–repressor species and we modeled the binding and unbinding reactions of each to the operator:

equation image
(1)
equation image
(2)
equation image
(3)

The stoichiometry of inducer–repressor binding is currently subject to debate [64]; it is unclear whether the affinity of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e081.jpg for the operator is of the same order as that of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e082.jpg or much lower. We therefore compared the effect on our model of both a high An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e083.jpg (comparable to An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e084.jpg) and a low An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e085.jpg (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e086.jpg). In either case, the affinity of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e087.jpg for the operator is thought to be low and we assumed An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e088.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e089.jpg. Values for the rate constants were obtained by fitting the model with experimental LacY distributions from single cells, as presented in Results.

Transcription, translation, and degradation

In the cell both transcription and translation are multistep processes involving numerous intermediates. Knowledge of the rate constants for each step of these processes in vivo is limited. We therefore assumed that each process was controlled by a single rate-limiting event and used pseudo first order rate equations with effective transcription and translation rates in the model. Such an approximation is reasonable as long as i) the concentration of the transcription and translation machinery is high and constant and ii) the non-Poissonian time delays missed by modeling such multi-step processes as a single step are not significant. For our model these conditions appear to be satisfied as the components of transcription and translation are among the most abundant in the cell and even though some, such as ribosomes, diffuse quite slowly there should nevertheless be a ready supply available at all times. Also, the times delays in these processes are on the order of seconds [27], while the dynamics of the cell response (here determined by the protein lifetime) is on the order of an hour.

Transcription of a LacY messenger RNA (mRNA) (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e090.jpg) from the lac operon was modeled as a first order process dependent on a free operator:

equation image
(4)

The effective transcription rate constant (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e092.jpg) was a free parameter determined during model fitting.

Decay of and translation from An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e093.jpg were modeled as a competition between RNase E enzymes [65] and ribosomes for an An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e094.jpg's ribosomal binding site (RBS). The rate of degradation of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e095.jpg by RNase E was chosen to result in a mean lifetime (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e096.jpg) of 90 s, as reported by Yu et al. [21]. The effective translation rate was chosen to produce a mean of four LacY proteins over the lifetime of an average An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e097.jpg messenger, also as reported in [21]:

equation image
(5)
equation image
(6)

with the effective rates given by An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e100.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e101.jpg.

The loss of membrane proteins in E. coli is primarily from dilution as a result of cellular growth over the cell cycle [66]. Therefore, degradation of LacY was modeled as a first order reaction with a half-life corresponding to the cell doubling time (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e102.jpg),

equation image
(7)

where An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e104.jpg.

Lac repressor/inducer kinetics

LacI rapidly dimerizes with very high affinity and the dimers further associate to form tetramers with a An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e105.jpg in the nanomolar range [67]. The tetrameric form enhances repression by binding multiple lac operators simultaneously [68]. However, modeling the DNA loops formed by this process would require additional inactive kinetic states, and as the focus of the current study is a two-state switch, we assumed a mutant form of LacI that did not tetramerize. This assumption allowed us to connect our model to single cell data from lac operator mutants incapable of DNA looping [22]. Furthermore, we assumed that the dimerization An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e106.jpg was sufficiently low that LacI only existed in the dimer state, the species An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e107.jpg. Ten molecules of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e108.jpg were placed in the cell, and we assumed the cell regulated this number to be constant, so that the noise from the transcription/translation of the repressor gene was ignored. Although noise from expression of LacI has been shown to have an effect on the induction rate [63], we chose to ignore this effect here to focus on the noise originating in the circuit itself.

The inducer molecules, isopropyl An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e109.jpg-D-1-thiogalactopyranoside (IPTG) and thiomethyl-An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e110.jpg-D-galactoside (TMG), are small sugar-like solutes that can either passively diffuse or be actively transported by LacY across the cellular membrane using an electrochemical proton gradient for energy. Inducer molecules in the extracellular space (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e111.jpg) and those in the intracellular space (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e112.jpg,An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e113.jpg) can diffuse across the membrane freely in both directions. The diffusive influx and efflux are modeled as first order reactions with the same kinetic constants,

equation image

where An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e115.jpg = An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e116.jpg An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e117.jpg [58], [69]. The active transport of inducer molecules into the cell from the extracellular space is modeled as an irreversible Michaelis-Menten reaction,

equation image

The value of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e119.jpg has been reported to be 12 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e120.jpg [70] and the Michaelis constant An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e121.jpg for the reaction with TMG to be ~500 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e122.jpg [71]. Similarly, the intracellular TMG concentration has been reported to be ~70-fold higher than the extracellular concentration in fully induced cells [63], [72]. Under pseudo steady state conditions, the ratio of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e123.jpg to An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e124.jpg is related to An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e125.jpg by the expression An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e126.jpg. In our model a maximal enrichment ratio of 70 corresponded to a An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e127.jpg of 400 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e128.jpg, which is the value we used for TMG. The active influx of inducer molecules through the membrane is given by the standard Michaelis-Menten expression An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e129.jpg, with An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e130.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e131.jpg. However, for simulations of stochastic kinetics, we need unique values for An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e132.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e133.jpg. As long as the relationship between An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e134.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e135.jpg corresponds to the required An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e136.jpg value, the influx of inducer will be correct irrespective of the values used. Only fluctuations in the external environment will be affected by the choice. In particular, a large value for An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e137.jpg relative to An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e138.jpg would require a correspondingly large value for An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e139.jpg, which would produce many unproductive inducer binding events outside the cell. These small fluctuations in the external inducer environment are expected to have little to no effect on gene expression dynamics inside the cell, but can have a significant impact on simulation cost. In the absence of any experimental data regarding the actual kinetic rates, we chose to make the An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e140.jpg rate be 1% of the An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e141.jpg rate to improve simulation performance by minimizing the calculation of such non-productive binding events. The value of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e142.jpg was then fixed by the relationship An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e143.jpg.

Upon entering the cell, inducer molecules can bind to free LacI and the repressor-operator complex, albeit with a much lower affinity. As each LacI monomer binds a single inducer molecule, there are three possible repressor dimer species, An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e144.jpg, An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e145.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e146.jpg, which interconvert according to the following reactions:

equation image
(8a)
equation image
(8b)

In vitro kinetic data suggests non-cooperative binding (Hill coefficient of 1) of inducer to An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e149.jpg in the absence of lac operator DNA [64], [73], [74], corresponding to An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e150.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e151.jpg (see Supporting Text S1).

Although there is some equilibrium data suggesting the binding to the complex is cooperative with a Hill coefficient of 1.45 [74], such cooperativity was not observed in kinetic measurements of binding and unbinding [75]. For simplicity, a non-cooperative model was assumed:

equation image
(9a)
equation image
(9b)

with An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e154.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e155.jpg.

Experimentally, different aspects of the lac circuit have been investigated using the inducers IPTG and TMG. To make use of the data one has to take into account the differences in their An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e156.jpg value, defined as the inducer concentration at which half of the LacI monomers are bound with an inducer. Kinetic and equilibrium binding measurements [74], [75] were available for IPTG binding to both free repressor and the repressor-operator complex. From the kinetic measurements, the rate constants for inducer binding and unbinding were An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e157.jpg An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e158.jpg and 0.2 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e159.jpg for the free repressor and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e160.jpg An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e161.jpg and 1.0 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e162.jpg for the repressor-operator complex. This yielded a An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e163.jpg for binding to the repressor-operator complex (89 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e164.jpg) that is ~20 times higher than for free repressor (4.1 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e165.jpg). Figure 3 shows the results of using these rate constants in stochastic simulations of inducer binding; good agreement between simulations and experiments were seen for both kinetic and equilibrium measurements. TMG has been reported to have a An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e166.jpg for binding to free repressor greater than that for IPTG by a factor of ~10 [76]. However, since neither kinetic data nor detailed equilibrium studies were available, we assumed the same unbinding rate constants for TMG as IPTG and left binding to both the free repressor and repressor-operator complex as free parameters in the model to be fitted from single molecule experimental data.

Figure 3
Fits of rate constants for IPTG binding to the lac repressor.

Well-stirred and spatially resolved simulations

Since the stochastic switch model is more complex than can be solved using analytic methods, we used computational Monte Carlo methods to sample the master equation and estimate the probability distributions. Two stochastic approaches were used to simulate the lac kinetic model: a well-stirred method using the CME and a spatially resolved method based on the reaction-diffusion master equation (RDME). The RDME model of the lac circuit can be thought of as a superset of the CME model in that all of the kinetic rates used for modeling reactions in the CME based model are also used in the RDME model, but with additional parameters regarding the spatial localization of particles and their diffusion in three-dimensional space.

Well-stirred model

The well-stirred model was sampled using a version of Gillespie's SSA algorithm [16] implemented in CUDA and running on the GPU. The lac model includes independent volumes for the extra- and intracellular space. These were tracked separately during the simulations and, to balance the flux of inducer across the membrane at equilibrium, the internal and external volumes were taken to be equal. Inducer in the extracellular space was maintained at a constant concentration for the propensity calculations by holding the number of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e169.jpg molecules fixed. Unless otherwise noted, we ran 10,000 replicates of each simulation to obtain sufficient sampling of the probability distributions.

Spatial model

The RDME model was sampled using our lattice microbe method. The multi-particle, in vivo diffusion operator for this method, based on that of Karapiperis and Blankleider [77], has been presented previously [37] and the reaction operator is provided in Supporting Text S1. Complete implementation details will be presented in a forthcoming publication. The method discretizes space onto a three-dimensional lattice of uniform spacing and time into uniform time steps. Chemical species randomly diffuse and react on the lattice according to rules defined using spatially dependent stochastic rate constants. A virtual microbe is constructed on a lattice by placing particles and obstacles on the lattice and specifying the reaction and diffusion properties of the lattice sites to mimic the spatial organization of the cell. The lattice microbe method uses the GPU as a computational coprocessor and the whole-cell simulations performed for this study were run for one hour of simulation time (slightly longer than a cell cycle of 55 min) using a time step of 50 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e170.jpg and a lattice spacing of 16 nm. For each external condition, 100 independent cell-scale simulations were run. Simulations were carried out on the NCSA Lincoln Intel 64 Tesla Cluster containing two NVIDIA Tesla GPU accelerators per node. Approximately 200 GPU-hours were required per hour of simulation time. Simulations of a smaller volume using 5 ns time steps and 2 nm lattice spacing were also performed to examine the variation in repressor rebinding as a function of packing density at the operator site.

Spatial models were developed for two E. coli phenotypes: fast- and slow-growth. For the fast-growth cells, the cellular volume was constrained to a typical E. coli cell shape: a cylinder with spherical end caps 2 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e171.jpg long by 0.8 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e172.jpg in diameter. The volume was surrounded by an impermeable cytoplasmic membrane separating the extracellular environment from the cytoplasm. We excluded the outer membrane from the model as its permeability is not thought to be a limiting factor for inducer transport. The intracellular space was randomly filled with stationary, in vivo obstacles to 50% volume fraction approximating the E. coli intracellular environment (see Table 2). The model was then coarse-grained onto a 16 nm resolution lattice, as described in Supporting Text S1.

Table 2
Obstacle abundance in in vivo spatial models.

For the slow-growth phenotype, we based the spatial model on cryoelectron tomography (CET) of an E. coli cell undergoing slow growth. The full cell was approximately 3 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e173.jpg long by 0.4 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e174.jpg wide and tomograms encompassed approximately one-third of the cell length. Ribosomes were matched and located in the tomograms as described previously [50] as was a portion of the cytoplasmic membrane. The position of the missing membrane was extrapolated to form a contiguous surface and ribosomes were placed at their measured positions in lieu of random placement. The completed one-third cell model was then mirrored to produce the opposite cell pole. The middle third was randomly generated using ribosome densities from the adjacent CET data. Ribosomes were not observed in the central volume of the cell, which we inferred to be the condensed nucleoid. A random walk algorithm was used to place a full-length E. coli chromosome in the nucleoid region. Starting from the center of the nucleoid region, cylinders representing DNA persistence lengths 50 nm in length and 2 nm in diameter were randomly added end-to-end such that the angle between successive cylinders was constrained to An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e175.jpg. If the random walk left the nucleoid region, the path was unwound a number of steps and a new random path started. This process was repeated until all 31,000 persistence lengths had been added. Since later coarse graining of the model onto a 16 nm resolution lattice spread out the nucleoid density, we did not constrain the chromosome to be circular in this model. Following nucleoid addition, the non-ribosomal in vivo obstacles were proportionally placed in the model at random locations (including within free space in the nucleoid region) to reach an occupied volume fraction of 50% (Table 2). The final slow growth model was then coarse-grained to a 16 nm lattice for simulation.

In E. coli, translation of an mRNA containing the sequence for an integral membrane protein is thought to be coupled with translocation of the resultant protein across the cytoplasmic membrane by the Sec translocase [78], i.e., cotranslational translocation. Specifically, LacY has been observed to require the bacterial signal recognition particle (SRP) pathway for functional membrane integration [79][81]. However, it is not presently clear whether or not transcription and translation of membrane proteins are also coupled such that the gene being transcribed is also physically located near the site of translocation. For this reason, we modeled two variants of operator placement for comparison: in the fast-growth phenotype the operator site was located in the center of the cell and in the slow-growth phenotype the operator site was placed on the nucleoid near the membrane (~32 nm away) close to a cell pole. These two different configuration allowed us to compare the mRNA dispersions for close versus far gene–translocation distances.

Messenger An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e176.jpg molecules were created at the location of the operator following transcription and then allowed to diffuse in the cytoplasm with a diffusion constant of 0.1 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e177.jpg [82], [83]; in the slow growth model An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e178.jpg molecules were precluded from entering the nucleoid. In the spatial model, then, An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e179.jpg was required to diffuse to the membrane before translation could occur; Equation 6 was limited to membrane sites. Since ribosomes likely attach to an mRNA's RBS while transcription is still ongoing [84], the model assumed that An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e180.jpg molecules were protected from degradation by RNase E until after An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e181.jpg reached the membrane; Equation 5 was also limited to membrane sites. Translation of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e182.jpg produced LacY proteins at the same location as the mRNA in the membrane. LacY molecules were constrained to diffuse in the membrane with a diffusion coefficient of 0.1 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e183.jpg.

An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e184.jpg molecules were randomly placed in the cell and diffused at 1 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e185.jpg within the intracellular volume. Small inducer molecules diffuse at ~1000 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e186.jpg in extracellular space and ~100 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e187.jpg in intracellular space. However, in order to reach simulation times on the order of the cell cycle the maximum diffusion coefficient in the model, which depends on the lattice spacing and time step, was 1.28 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e188.jpg. Therefore, the diffusion coefficient of the inducer molecules was set to 1.28 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e189.jpg. Since inducer molecules are present in large numbers and they diffuse faster than the repressor this approximation was not expected to have a noticeable effect. The lattice was connected to a infinite reservoir of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e190.jpg molecules through the use of constant concentration boundary conditions to maintain the extracellular space at a constant inducer concentration.

Maximum likelihood fitting of gene expression models

We analyzed the capability of the burst and two-state analytic models of gene expression to recover parameters from our stochastic simulations of an inducible switch by fitting molecular distributions. We used a maximum likelihood method to estimate the model parameters. Briefly, the likelihood An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e191.jpg of the model parameters An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e192.jpg having produced a set of observations An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e193.jpg is given by

equation image

where An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e195.jpg is the conditional probability of observation An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e196.jpg occurring given the parameters An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e197.jpg. The parameters that maximize this likelihood function are those that describe the best fit of the model to the data, assuming a uniform prior distribution for the parameter probabilities. To find the best parameters for a model of gene expression, An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e198.jpg was calculated using the model's steady-state probability density function with the An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e199.jpg values being the protein counts from the 10,000 simulations. The parameter values that minimized the negative log of the likelihood function were then found using downhill simplex minimization as implement in the Matlab fminsearch function. We estimated the confidence intervals for different sample sizes by taking 1000 random sets of either 50 or 200 cells from the full set of 10,000 and performed maximum likelihood estimation on each of these data sets. The confidence range for each parameter was then defined by the middle 95% of the values obtained during these random resamplings.

The burst model was first expressed in terms of parameters An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e200.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e201.jpg by Friedman et al. [24] as the Gamma distribution. However, since our stochastic simulations produced discrete protein counts, we used the discrete formulation for the steady-state probability density derived by Shahrezaei and Swain [25] in terms of a negative binomial distribution

equation image
(10)

with parameter An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e203.jpg being the burst frequency (bursts per mean protein lifetime) and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e204.jpg being the burst size (proteins produced per burst).

The two-state model was fit using the steady-state probability density function derived by Shahrezaei and Swain [25]:

equation image
(11)
equation image
(12)

In this expression the parameters are An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e207.jpg, An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e208.jpg, An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e209.jpg (the activation rate), and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e210.jpg (the inactivation rate), the latter two being expressed in units of mean protein lifetime. Additionally, An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e211.jpg, An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e212.jpg, An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e213.jpg, and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e214.jpg is Gauss's hypergeometric function. Fitting with all four parameters free often resulted in convergence in a local minima, so we adopted a fitting procedure whereby we first constrained the An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e215.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e216.jpg parameters and fit only An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e217.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e218.jpg to obtain initial estimates of these two parameters. In the fully induced state An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e219.jpg the above probability density function reduces to a negative binomial distribution with no dependence on An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e220.jpg or An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e221.jpg, only An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e222.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e223.jpg. Since neither An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e224.jpg nor An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e225.jpg depend on inducer concentration, it is a reasonable approximation to use the values for An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e226.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e227.jpg in the fully induced state as initial estimates for all inducer concentrations. After obtaining an initial fit for An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e228.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e229.jpg, we then performed another fit with An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e230.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e231.jpg unconstrained and with An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e232.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e233.jpg allowed to vary ±5%. This procedure resulted in convergence at a higher likelihood score than when all four parameter were fit simultaneously for all distributions except one.

Results

Here we present the result of our study into the noise effects in the inducible lac genetic switch. The first two sections describe the fitting of model free parameters to data from single-molecule fluorescence studies on E. coli populations. The next two sections analyze noise in the well-stirred circuit due to its regulatory control elements. The final two sections report on changes to the behavior of the circuit from in vivo effects, using a model of a spatially heterogeneous, crowded cell and then an experimentally determined cell structure under an alternate growth phenotype.

Linear relationship between transcriptional burst size and inducer concentration

In a recent in vivo single-molecule fluorescence study, Choi et al. measured the distributions of a fluorescent reporter protein under control of the lac operator in individual E. coli cells at various inducer (TMG) concentrations [22]. They performed the measurements in the absence of LacY's positive feedback by replacing its gene with that of the membrane protein Tsr in the lac operon. This enabled an accurate determination of the protein distribution produced by the circuit at a given inducer concentration without any confounding non-linear effects due to enhancement of the internal inducer concentration by LacY. In the absence of DNA looping, they were able to fit their observed distributions to a gamma distribution An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e234.jpg, where An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e235.jpg was interpreted as the frequency of transcriptional bursts relative to the protein lifetime and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e236.jpg as the mean number of proteins produced per burst. They observed a relatively constant value for the burst frequency of 3–4 and a linearly increasing relationship between burst size and inducer concentration at low to intermediates concentrations.

To understand the origin of the linear relationship between burst size and inducer concentration and to reproduce this behavior in our model, we derived an expression for the burst size as a function of kinetic parameters in our model. As long as bursts are infrequent relative to protein degradation, i.e. once a free operator is bound with a repressor it remains bound for a significant fraction of the cell cycle, transcriptional bursting from the lac operon can be modeled as a Markov process with competition between RNA polymerase (RNAP) and the various LacI species for binding to the free operator (see Figure 4).

Figure 4
Markov diagram for transcriptional bursting in the lac circuit.

Transcription initiation by RNAP was modeled as a pseudo first order process (Equation 4), with a rate constant of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e239.jpg. The two repressor states with potentially significant binding affinity were An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e240.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e241.jpg, shown in Equations 1 & 2. Free repressor binds with free operator with a rate constant of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e242.jpg resulting in a pseudo first order rate of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e243.jpg. Given the current debate surrounding the binding affinity of the An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e244.jpg state to the operator, we set the rate constant An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e245.jpg to be proportional to the free repressor binding constant and analyzed the effect of varying the proportionality constant An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e246.jpg on the pseudo first order rate An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e247.jpg. This model of transcriptional bursting assumes that the binding of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e248.jpg to the free operator is negligible at low inducer concentrations by assuming An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e249.jpg and ignoring Equation 3. In practice, this condition was satisfied when An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e250.jpg. We used the upper limit An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e251.jpg in our model, which is within the range experimentally reported [76].

Following the unbinding of a repressor from the repressor–operator complex, the probability of transcription initiation (and subsequent mRNA creation) occurring at the free operator as opposed to a repressor re-binding is

equation image
(13)

The probability of a given number of consecutive transcription initiation events (the size of the mRNA burst) then follows a geometric distribution with An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e253.jpg of which the mean is An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e254.jpg. However, repressor unbinding events that produce no mRNA are not observable as a burst, therefore the mean number of mRNA produced in a transcription bursts (B) is

equation image
(14)

Combining Equations 13 and 14 gives the expression for the mean transcription burst size in terms of the rate constants for transcription initiation and repressor binding

equation image
(15)

Given the inducer mass balances (see Supporting Text S1) and the expression for the total number of repressor dimers An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e257.jpg, one can derive the equilibrium concentrations of the two repressor species

equation image

where An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e259.jpg is the inducer concentration at which half of the repressor monomers are bound to an inducer molecule. Substituting An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e260.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e261.jpg into Equation 17 gives the expression for the transcription burst size as a function of inducer concentration

equation image
(16)

From this last equation it is clear that the transcription burst size will be linear over the entire range of inducer concentrations only when An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e263.jpg. Figure 5 shows the effect of varying An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e264.jpg, of particular interest are the very low values. When An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e265.jpg, the transcription burst size does not linearly increase over the range of inducer concentrations for which this behavior has been reported (0–200 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e266.jpg). In the model here formulated, a linear relationship between size and inducer concentration exists only when the binding affinity of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e267.jpg for the free operator is comparable to that of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e268.jpg. For our simulations, we therefore chose An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e269.jpg, such that An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e270.jpg, as this value assumed no effect on the unbound repressor monomer due to a single bound inducer and gave a strictly linear relationship for all inducer concentrations.

Figure 5
Parameter space of repressor binding parameter An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e271.jpg.

Fitting transcription and inducer/repressor rate constants to single-cell distributions

To obtain values for the model parameters An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e284.jpg, An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e285.jpg, and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e286.jpg, we used the distributions for LacY reported by Choi et al. [22], specifically the inferred burst frequency (bursts per cell cycle) and size parameters (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e287.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e288.jpg) from their gamma distribution fits. From Equation 18, the mean transcription burst size as a function of inducer concentration is An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e289.jpg. This equation is linear in inducer concentration and by fitting it (multiplied by the mean number of proteins produced per mRNA) to the experimental protein burst sizes, as shown in Figure 6, one can constrain the kinetic parameters. The y-intercept of the line fixes the ratio of transcription to repression in the uninduced state (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e290.jpg) and the slope can then be used to obtain An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e291.jpg = 17.6 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e292.jpg for TMG.

Figure 6
Linear fit of burst size to inducer concentration.

The linear fit, however, only fixes the ratio between An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e293.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e294.jpg. To recover unique values for these two rate constants, we next considered the mean duration of each transcription burst. The interpretation of the shape parameter An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e295.jpg of the gamma distribution as the burst frequency is only meaningful if the burst duration is short compared to the protein lifetime. In that case, individual exponentially sized bursts can be considered exponentially distributed in time and therefore act independently to give rise to a gamma distribution of protein abundance. In setting rate constants for the model, then, we wanted to ensure that the burst duration was appropriately short.

The burst duration An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e296.jpg is simply the mean time for a repressor to bind to a free operator. Given a constant An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e297.jpg, a linear relationship between burst size and inducer concentration also implies a linear relationship between An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e298.jpg and inducer concentration as can be seen from

equation image
(17)

where An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e300.jpg in the last step. For TMG, the linear relationship between burst size and inducer concentration extended to at least ~200 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e301.jpg, which is ~11 times the An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e302.jpg value for TMG of 17.6 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e303.jpg. From Figure 7 it can be seen that the interpretation of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e304.jpg as the burst frequency begins to break down once An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e305.jpg is >5% of the protein lifetime. Using 5% of the protein lifetime as An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e306.jpg for 200 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e307.jpg, we can compute the value for An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e308.jpg that gives the appropriate An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e309.jpg: An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e310.jpg An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e311.jpg An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e312.jpg, using a cell doubling time An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e313.jpg of 55 minutes. With this value for the repressor binding rate, a single repressor molecule in an E. coli cell would take ~200 s to find a free operator. This is somewhat faster than the 354 s reported by Elf et al. [51]. Using the above value for An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e314.jpg and the ratio of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e315.jpg to An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e316.jpg from the linear fit of the experimental data we obtained the value for the transcription rate An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e317.jpg = An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e318.jpg An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e319.jpg. This rate for transcription initiation resulted in a steady state concentration of ~2500 LacY molecules per cell in the fully induced state, within a factor of two of the ~1000–1200 reported in the literature [22], [26]. The value also falls within the range of 1000–3000 seen for other highly expressed proteins in E. coli [85]. Accurate measurements of the burst duration in the lac system, particularly in the fully induced state, would increase the accuracy of our model.

Figure 7
Burst analysis of stochastic simulations of a simple two-state process.

In order to reproduce a burst frequency of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e331.jpg over the mean LacY lifetime in the model, the repressor should dissociate from the operator with a frequency An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e332.jpg, assuming that each dissociation event produces a burst and that An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e333.jpg An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e334.jpg the cell cycle. The burst frequencies inferred by Choi et al. for TMG levels ≤100 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e335.jpg are relatively constant with a mean of ~3 bursts. This corresponds to An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e336.jpg = An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e337.jpg An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e338.jpg. Since the dissociation of a repressor dimer is not thought to be significantly affected by the binding of a single inducer molecule, An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e339.jpg = An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e340.jpg. The affinity of a repressor dimer with two bound inducer molecules, however, is thought to be much lower, i.e., the binding of a second inducer molecule essentially knocks the repressor off of the operator. In the absence of this effect, the response to an increase in inducer concentrations would take a significant fraction of the cell cycle. Elf et al. reported a response time of <60 seconds for addition of IPTG to concentrations from 50 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e341.jpg – 1 mM [51]. Therefore, we fit An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e342.jpg such that the response of the model to increase in IPTG agreed with the published data. The best fit value was obtained for An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e343.jpg (shown in Figure 8A).

Figure 8
Parameter fitting for inducer–repressor–operator interactions.

The final kinetic rates to be defined were those regarding the binding of TMG to the repressor–operator complex (Equations 10 & 11). As discussed in Methods, we used the same dissociation rates as for IPTG, leaving only the association rates An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e346.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e347.jpg, both of which can be derived from the An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e348.jpg value, which is the inducer concentration at which half of the repressor–operator complexes have a bound inducer. Figure 8B shows the effect of varying An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e349.jpg on the burst frequency. As An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e350.jpg approaches An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e351.jpg, the burst frequency begins to diverge from its expected value. This is due to the increasing occupancy of the An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e352.jpgO state, which can decay much more quickly into a free operator than the other repressed states; with operator free more often, there are more bursts over the lifetime of a protein. A value of 3 mM for An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e353.jpg gave the best agreement with the experimental burst frequencies for TMG.

Population distributions without positive feedback

Using the derived rates, we performed well-stirred stochastic simulations of the lac model in the absence of LacY positive feedback (NPF model), obtaining the stationary LacY distributions as a function of internal inducer concentration shown in Figure 9. Compared to the intrinsic noise of the two-state model, the NPF model contains additional noise contributions from the non-constant rates for transitioning between active and inactive transcriptional states. The distributions showed the widest cell-to-cell variability due to the intrinsic noise of the system at intermediate inducer concentrations of 50–400 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e354.jpg. At high inducer concentrations the population migrated toward a less variable distribution, as expected. Up to 100 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e355.jpg, the population distributions agreed well with those reported by Choi et al. but at 200 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e356.jpg the agreement began to break down. This discrepancy at concentrations >100 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e357.jpg was caused by two primary factors: the burst duration and the action of inducer knocking repressor off of the operator. Increasing the repressor binding rate would improve the fit by decreasing the duration of each burst, but would cause a large increase in the total number of LacY molecules in the fully induced state, which is not supported experimentally. Alternatively, one could increase the An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e358.jpg value, causing less inducer instigated dissociation of the repressor–operator complex, but this would decrease the responsiveness of the circuit to addition of inducer, which is also not supported experimentally. Clearly, in order for the model to have greater predictive power, additional features would be necessary. For example, adding a delay between production of mRNA to account for the steps of RNAP open complex formation or more detailed modeling of translation. But lacking the in vivo experimental results to validate any additional complexity, we chose to ignore these effects and analyzed the model as described.

Figure 9
Steady state LacY distributions from the well-stirred NPF model.

The gene regulation function (GRF) of an genetic system describes the relation between the activity of a gene and its regulatory control elements [86][88]. In the steady state, protein production is balanced by protein degradation/dilution. The mean protein count as a function of the control elements provides a method to analyze a GRF. The mean number of LacY per cell as function of the TMG concentration (Figure 9D) and the fraction of time spent in the transcriptionally active state (Figure 9F) show the regulatory behavior of the NPF model. We saw a typical sigmoidal regulatory response that was well fit by a Hill equation with an inflection at 312 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e360.jpg and a Hill coefficient of 2.11. In a stochastic system, though, the mean rate of gene expression is just one piece of information. As important for a stochastic GRF is how the distribution changes with inducer concentration. The Fano factor (variance/mean) provides a measure of the variation of the distribution. For reference, the Fano factor of a Poisson process is 1. For the NPF model (Figure 9E) the Fano factor monotonically increases until 100–200 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e361.jpg where it peaks at a value of ~60 and then begins to decrease ending at a lower value of relative noise than at zero inducer.

Noise due to positive feedback

Next we investigated noise in the inducible genetic switch when the positive feedback regulatory link was active (PFB model). The lacY gene located in the lac operon codes for the integral membrane protein LacY, which actively imports inducer molecules (lactose/An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e362.jpg co-transport) establishing a positive feedback loop as shown in Figure 1C. The presence of active LacY in the membrane creates a concentration gradient enriching the intracellular environment with inducer molecules relative to extracellular space. For a fixed An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e363.jpg concentration, the underlying GRF for the lac operon therefore operates not only at an increased inducer concentration but, since the number of LacY is different for each cell, across a distribution of internal inducer concentrations.

We calculated the population distributions for the PFB model using well-stirred stochastic simulations at various An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e364.jpg concentrations. Starting from a stationary population distribution in the absence of inducer, each population of 10,000 cells was subject to an instantaneous increase in An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e365.jpg and simulated for twenty-four hours. Above an An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e366.jpg concentration of ~10 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e367.jpg, cells in the population began to switch to an induced state in which LacY expression was near its maximum value (see Figure 10A and B). Above ~25 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e368.jpg the transition to full expression was relatively concerted throughout the population. In the range of 10–25 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e369.jpg, though, there were two transiently stable subpopulations, one uninduced and the other induced – the overall population was bimodal for a time.

Figure 10
Response of an uninduced PFB population to the addition of external inducer.

To quantify the switching behavior of the population, we classified cells at regular time intervals as uninduced with <400–600 LacY (best fit for each An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e374.jpg) or induced with >1750 LacY. Each subpopulation was then analyzed separately. The mean and variance of the distributions (Figure 10C) show that, after an initial response phase, the distribution of the uninduced subpopulation was stable over time. This was true even as the total number of cells in the uninduced population was decreasing as cells within it were switching to the induced state. At intermediate inducer concentrations, the uninduced cell population appeared to reach a stationary distribution from which cells independently and stochastically transitioned to the induced state. In contrast, at higher inducer concentrations the population migrated as a whole in a more downhill-like manner.

Noise in a GRF can be expressed in terms of its effect on the phenotypic variance in a population under identical environmental conditions. To compare noise between the NPF and PFB models, we first mapped An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e375.jpg concentrations to mean An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e376.jpg concentrations in the uninduced and induced subpopulations (in the NPF model An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e377.jpg = An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e378.jpg). We then compared both the mean of the LacY distributions and the Fano factor for the two models. The mean values for the LacY distributions (Figure 11) were similar but the noise in the uninduced subpopulation was significantly higher in the model with positive feedback. Since the underlying GRF is equivalent between the two models, it is the action of the GRF on the distribution of An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e379.jpg concentrations that gives rise to the increase in intrinsic noise in the PFB model.

Figure 11
Effect of positive feedback on GRF noise.

Differences in circuit behavior due to in vivo crowding

Having established the well-stirred PFB stationary distribution, we next evaluated the effect of in vivo molecular crowding on the distributions, the PFB+IV model. One obvious reaction subject to spatial effects is the rebinding of the repressor to the operator following an unbinding event. Immediately after unbinding, a repressor is necessarily localized near the operator, i.e. it has a memory of its location. As was shown by van Zon et al. [27], this memory effect increases the probability of repressor rebinding at very short times compared to a well-stirred approximation. Previous studies only considered the effect of normal diffusion following unbinding but there is an additional effect caused by anomalous diffusion due to in vivo crowding.

To investigate repressor rebinding in an in vivo environment, we performed reaction-diffusion simulations of a An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e380.jpg volume centered on an operator immediately following unbinding of a repressor. We varied the packing density of the approximated in vivo environment to study its effect on rebinding. Figure 12A and B shows that there is an anomalous effect at short time scales (<1 ms). Repressor diffusion at very short time scales is normal at the in vitro rate, but between 1–100 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e381.jpg there is a period of anomalous behavior, and at very long time scales repressor diffusion returns to normal diffusion behavior with a lower diffusion coefficient D. Brownian dynamics simulations of proteins in a virtual in vivo environment [39] show a similar anomalous behavior when including only steric constraints with a minimum in the time exponent of ~0.8 for proteins slightly larger than the 75 kDa repressor dimer. When electrostatic effects are included in the Brownian dynamics simulations, however, the apparent diffusion coefficient as well as the anomalous exponent change greatly, so our results should only be considered an upper bound on the in vivo effects. Including further electrostatically driven interactions such as non-specific binding, will increase the anomalous behavior of the repressor.

Figure 12
The effect of in vivo crowding on repressor rebinding.

The anomalous behavior of the repressor causes it to spend more time near the operator following unbinding than would be expected for purely Brownian diffusion, leading to more encounters with the operator and a potentially greater probability of rebinding. To measure the change in rebinding probability, we counted the number of repressors that rebound to the operator following unbinding versus the number that escaped into bulk solution, defined here as leaving the simulation volume. As can be seen in Figure 12C, as the density of in vivo crowding increases, the probability of rebinding goes up. Compared to an in vitro unpacked environment at 15% probability of rebinding, at 50% packing the probability of rebinding is ~24%. The distribution of escape times also broadens (Figure 12D) with particles in general taking longer to diffuse away. The anomalous memory effect resulted in the duration of some bursts being significantly shorter than expected.

To study the effect of burst duration differences on the stationary LacY distributions in a population, we used our lattice microbe method to generate PFB+IV trajectories of spatially resolved rapid-growth E. coli cells (see Methods). Beginning with the stationary distribution from the well-stirred PFB population, 100 cells were simulated at five internal inducer concentrations for one hour, slightly longer than the duration of a cell cycle (55 minutes), see Video S1. Over the course of the simulations, distributions in the in vivo models gradually migrated to lower mean values and lower noise, as can be seen in Figure 13. Two factors caused this migration: First, the shorter burst durations due to the anomalous diffusion effect described above resulted in fewer proteins being produced per burst and more time spent in the inactive state led to more frequent bursts and less noise. Second, the effective increase in repressor due to the decreased reaction volume. In contrast to spatial effects in an in vitro environment [27], it appears that in vivo crowding lowers both the mean value and the noise in distributions of observables. Since bacterial cells such as E. coli are known to have packing density changes during different portions of the cell cycle and/or growth conditions, this presents the possibility of measuring these in vivo effects on living cells if the observable distributions can be accurately quantified as a function of the cell cycle or growth conditions.

Figure 13
LacY PFB+IV in vivo distributions.

Whole-cell modeling using experimentally determined cell architecture

As a first attempt at addressing how changes in the cellular environment due to growth conditions affect gene expression noise, we used CET of E. coli cells under slow growth to build a whole-cell model of an individual bacteria (Figure 14A). Under conditions of slow growth in minimal media E. coli B/r K grows as elongated cylinders with diameter ~400 nm [89], which are amenable for CET [50]. The tomograms were used to identify the membrane-enclosed volume of an individual cell along with the three-dimensional position of ribosomes within it. The E. coli B/r K cell under slow growth had only An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e387.jpg of the volume of typical fast growing cells. A central region of the cell was devoid of ribosomes and inferred to be the location of the condensed nucleoid.

Figure 14
Analysis of cryoelectron tomography based cell model.

We studied the operation of the lac circuit in the slow-growth phenotype (PFB+IV+CET) using 100 random replica cells. Each replica used the same experimentally measured cellular geometry and ribosome positions, but a random distribution of other molecules including a condensed chromosome (see Methods for details). Cells were simulated using the lattice microbe method in 15 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e388.jpg external TMG, starting with LacY and mRNA counts sampled from the uninduced stationary distribution of the well-stirred PFB model, see Video S2. Simulations were run for either one hour or until the cell had induced, whichever came first. There were clear differences between the slow- and fast-growth in vivo models. Of the 100 slow-growth cells, 11 induced within one hour whereas only a single fast-growth cell induced in the same time period. Also, the mean number of LacY molecules in the uninduced slow-growth population increased ~15% over the course of one hour, compared to the fast-growth population which decreased ~15%.

Analysis of the simulation trajectories revealed that the primary cause of the differences in LacY distributions between the slow- and fast-growth models was an increased mean inducer concentration in the smaller cells, 100 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e389.jpg versus 42 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e390.jpg. For a given number of LacY proteins, the cells with the smaller volume had an increased internal inducer concentration. The increased levels of inducer caused a slight lengthening of the mean duration of free operator events, 68 seconds versus 64 seconds, and a corresponding larger burst size. A bigger change was observed in the mean lifetime of the repressor–operator complex, which decreased to 430 seconds from 730 seconds (Figure 14B,C). The decrease effected an increase in the mean number of transcription bursts per hour, to 4.3 from 2.6.

The slow-growth model provides a first approximation as to the effect of differences in cellular architecture on stochastic gene expression. The model assumed the same number of repressor molecules for smaller cells, which may not be accurate as repressor is known to regulate its own expression. However, since the largest effect was due to an increased rate of repressor unbinding due to elevated inducer levels, which is independent of repressor concentration, we consider the general result of increased burst frequency and rate of induction in smaller cells to be intriguing. It implies that there might be a difference in the switching properties during the first part of the cell cycle following division when a large burst of LacY would have an increased influence on switching due to the reduced cellular volume. Such an effect could potentially be measured using cell synchronization techniques. Although specific ribosome placement likely also influenced repressor rebinding in the slow-growth model, any differences were overshadowed by the effect of the cell volume change. Nevertheless, in a situation where the placed macromolecules are involved in the reaction kinetics, we anticipate accurate (non-uniform) placement will take on much greater importance.

Another large difference between the slow- and fast-growth models arose due to the presence of a condensed nucleoid coupled with the smaller cell diameter. In the fast-growth cells the chromosome was assumed to be diffuse and not an obstacle for mRNA diffusion. In the slow-growth cells, the chromosome was randomly placed in the ribosome-excluded region observed in the tomograms and it represented an obstruction for mRNA diffusion. Additionally, the operator was positioned in the center of the fast-growth cells and at the edge of the nucleoid in the slow-growth cells. As can be seen in Figure 14D there was a dramatic increase in localization of mRNA in the slow-growth cells as a result of this arrangement. A recent report of mRNA localization in bacteria [90] suggests that the relative locations of transcription and translation in bacteria may indeed be correlated. If that is generally true, then in systems where the location of protein synthesis affects the reaction kinetics it will be important to know the actual position of the gene in the cell and measurement of the dispersion of the transcripts might be one way to quantify whether the gene is physically located near the site of translation and translocation.

Discussion

Fitting population distributions to models of stochastic gene expression

Fitting protein population distributions to gene expression models will be a key step in developing simulations of other stochastic cellular systems with predictive power. Parameters obtained from fitting the distributions will drive the computations. Our stochastic simulations of the inducible lac switch provide an opportunity to test the process of extracting parameters from a population distribution arising from a complex gene expression system using simplified but analytically tractable models. To do so, we fit the stationary population distributions from our simulations to both the burst and two-state models (Figure 1A & B) and evaluated their capability to recover the stochastic rate constants used in the simulations (e.g. An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e391.jpg, An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e392.jpg, etc). The analysis was performed for each of the different noise variations described above, corresponding to the NPF, PFB, and PFB+IV simulations. We excluded the PFB+IV+CET simulations from this study as they were not performed over a range of inducer conditions. The best fit parameter values were obtained by maximum likelihood estimation using the stationary probability density function (PDF) for the burst and two-state models, Equations 12 & 14 in Methods. Fits were performed using 10,000 cells for NPF and PFB simulations and 100 cells for PFB+IV simulations.

Figure 15A & B show parameter estimates obtained from fitting using the burst model's gamma distribution PDF (Equation 12). The An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e393.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e394.jpg parameters (the B subscript indicates parameters for the burst model) reliably recover the burst frequency and burst size, respectively, in the NPF simulations at low inducer concentrations, but diverge from the simulation values above ~100 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e395.jpg. This is as expected as the model is only valid when the duration of each burst is short enough that sequential bursts can be considered as occurring independently, <5% of the protein lifetime as shown in Results. In particular the divergence occurs near the switching threshold, making this model most suitable for analyzing the system in the uninduced state with low expression levels. However, the clearness of the biological interpretation for the model parameters as the burst frequency and size make the model extremely valuable over the regime it is valid.

Figure 15
Maximum-likelihood fitting of two models for gene expression to stochastic simulations of an inducible genetic circuit.

Fitting the NPF simulation data to the stationary PDF of the two-state model (Figure 1B; Equation 14) provides good parameter estimates over a wider range of inducer concentrations. The fits are shown in Figure 15C–F for the parameters An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e396.jpg (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e397.jpg; the TS subscript indicates two-state), An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e398.jpg (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e399.jpg), An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e400.jpg (the rate constant for operator activation), and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e401.jpg (the rate constant for operator inactivation), respectively. As the inducer concentration increases, though, many more cells are required to obtain reliable estimates. Using even 10,000 cells, we were unable to obtain good fits for the highest expression levels. At these inducer levels so little time is spent in the inactive state that the difference in likelihood values for different switching rates is insufficient to find a unique maximum using 10,000 samples. However, as the time spent in the inactive state approaches zero (An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e402.jpg) the probability distribution approaches a negative binomial distribution without dependence on An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e403.jpg or An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e404.jpg, so it is possible to estimate the An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e405.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e406.jpg parameters in the fully induced state by fitting to a negative binomial.

The two-state model therefore appears to be a reasonable method for fitting the NPF simulations. Using the fitting parameters (along with known or estimated mRNA and protein degradation rates), one can readily recover the transcription and translation rates as well as the rates of the operator switching between active and inactive states at a given inducer concentration. Even though switching between active and inactive states in the lac switch is not a first order process – it is controlled by 14 reactions – at a given inducer concentration the steady state switching times are reasonably well-approximated by a single exponential. A further improvement in the two-state model would allow An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e407.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e408.jpg to depend on the inducer concentration using, e.g., a Hill function. An analytic solution to such a model would allow extraction of parameters from a multivariate fit using data across all inducer concentrations. However, to the best of our knowledge, the analytic form of such a model has not been derived.

Using the steady state distributions from the PFB simulations, neither model achieves good fits. For the two-state model, the An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e409.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e410.jpg parameters are recovered correctly, but the fits for the An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e411.jpg and An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e412.jpg parameters are lower than expected. The poor fit for these parameters is due to noise in the switching rates of the cell population caused by differences in internal inducer concentrations. With positive feedback, it will be very difficult to reliably estimate model parameters from population distributions due to non-linear noise. Fitting to experimental data should be done in the absence of positive feedback, such as by using gene knock-outs to eliminate circuit components responsible for positive feedback. However, if an analytic model were developed including positive feedback effects, comparison of systems with and without these effects could provide estimates of positive feedback parameters, e.g. inducer transport rates.

Fits to the PFB+IV simulations as well show deviations from the expected values; in vivo crowding noise changes the parameter fits. For these simulations, an additional source of discrepancies with the models is the non-Poissonian behavior of repressor rebinding – there is a positional memory in the system for a short time following unbinding. In our simulations the effect from in vivo conditions due to excluded volume is modest, but there are other in vivo factors still not accounted for in them, especially non-specific binding as recently reported by McGuffee and Elcock [39], which would have an even larger effect on repressor rebinding. Also, repressor rebinding most likely occurs via a series of 1D sliding and 3D hopping steps, the effect of which on rebinding in a crowded environment is not known. Accounting for in vivo effects when deriving parameter from experimental population distributions, which would include in vivo noise contributions, will be difficult. Possibly an iterative process of refinement may be required, starting with model estimates and proceeding through multiple rounds of spatial simulation.

Overall, it appears that fitting population distributions to the two-state model could prove to be an effective way of obtaining rate constants for stochastic simulations of gene regulation. Single-molecule in vivo fluorescence imaging provides a way to experimentally measure these distributions. Measurements over a range of regulatory conditions could then be used to build a stochastic gene regulation function, provided the actual probability distributions from single-molecule experiments were available at each condition. However, it is important to acknowledge that our simulations did not include a contribution from global extrinsic noise. Noise in our simulations under conditions of high expression approaches Poissonian, as expected from the intrinsic noise of an uncorrelated random process. A recent study has clearly shown, though, that there is a constant level of global extrinsic noise in gene expression in E. coli, maintaining population heterogeneity even at high levels of gene expression [85]. This implies that a way to correct for the global extrinsic noise will be needed in order to fit experimental population distributions at high expression levels.

Probability landscape of an inducible lac switch

The probability distribution for a stochastic biochemical system to be in a particular state represents the totality of information about the system. From it various measures of the behavior of the system such as the mean first passage time between two states or their relative population at the steady state can be obtained. For models of stochastic gene expression, two relevant reaction coordinates are the number of protein and mRNA molecules in the system. We used our stochastic simulations to reconstruct the two-dimensional probability landscapes (negative log of the PDF) of the NPF and PFB models at two external TMG concentrations (Figure 16).

Figure 16
Probability landscape of protein–mRNA abundances in the inducible lac switch model.

The steady-state landscape of the NPF simulations at 500 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e415.jpg inducer shows a bistable mRNA distribution that has been reported by others [25], [91]. One minima is located near 0 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e416.jpg/1700 LacY and the other near 10 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e417.jpg/1800 LacY. Note that the stable 10/1800 point does not imply that 180 LacY were produced per An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e418.jpg, as the degradation rates of the two molecules differ. At 500 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e419.jpg inducer, the net time some cells stochastically spend in the inactive state is greater than the typical lifetime of the mRNA bursts. These cells then drift to a zero mRNA abundance. The higher density is caused by an accumulation of these cells near the zero mRNA level until their next mRNA burst pushes them back into a random cycle around the mean mRNA burst size. Interestingly, though, the protein distributions at the two mRNA minimum are different, with the protein abundance being slightly lower in the lower mRNA minimum. This means that the protein and mRNA probability distributions are not completely independent of each other; the joint probability distribution has cross terms. While not a large difference, it is nevertheless possible that the joint protein–mRNA distribution could be used to obtain better parameter fits for the two-state model with fewer cells, if the mRNA counts were known.

The bimodal distributions seen in the LacY distributions from the PFB simulations (Figure 10) are recapitulated in the probability landscape for switching. The two-dimensional landscape allows classification of both the uninduced and induced states in terms of their relative protein and mRNA abundances. Additionally, the landscape reveals the transition path for switching from the uninduced to the induced state. One can imagine two possible scenarios for the transition, either the gradual build-up of protein by a series of small bursts, or alternatively, by the random occurrence of a small number of larger bursts. For the lac system with DNA looping, Choi et al. [26] have persuasively argued for the random large bursts as the switching initiator. The switching mechanism of the stochastic system in the absence of looping, however, is not so clear. The probability landscape of our lac model suggests that it is actually the occurrence of several large mRNA bursts, on the order of >10 molecules, in quick succession that is responsible for putting the cell on the path to induction. Cells can spend a significant amount of time in a high LacY but low An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e420.jpg state without inducing. This behavior is apparent in the cell trajectory plotted in Figure 16B. Switching therefore is a process in which not only a protein threshold must be crossed, but also an mRNA threshold.

Conclusions

Our goal with this study was to go beyond previous stochastic simulations of the lac circuit by using information from single molecule protein distributions and experimentally determined cellular architecture to constrain the kinetic parameters and estimate the effect of spatial heterogeneity on the response of the switch. The kinetic model of the inducible lac genetic switch presented in this study illustrates the utility of incorporating single-molecule, single-cell data when modeling cellular biochemical systems. The model was derived using a kinetic framework reproducing a linear relationship between protein burst size and inducer concentration at low concentrations, as has been reported experimentally. Analysis of the linear relationship in terms of inducer–repressor–operator interactions suggests that the stoichiometry of repressor binding is such that repressor dimers with one bound inducer still have significant affinity for the lac operator. Furthermore, single-cell population distributions were used to obtain estimates of the effective rate constants for transcription and repression in the cell. With future increases in performance of the lattice microbe simulation method it should be possible to iteratively refine the kinetic rate constants to account for the effects of cellular architecture, such as we obtained here from CET experiments, and cytoplasmic crowding. Using such in vivo adjusted rate constants the in vivo models should then more accurately reproduce experimental population distributions, which are after all measured under in vivo conditions, than the well-stirred models.

The lac model without positive feedback provided a baseline for the noise in the regulation of the lac operon. Intrinsic noise at low gene expression was significantly higher than Poissonian and peaked when the promoter was active 10–30% of the time. The model with positive feedback produced similar mean values for a given intracellular inducer concentration, but the noise was substantially greater. We attribute this effect to the non-linear gene regulatory function operating on a distribution of intracellular inducer levels. Global extrinsic noise in the transcription/translation machinery is a large contributor to population heterogeneity at high levels of expression, but we excluded such noise from the current study.

Fitting of data from stochastic simulations of the lac switch with the burst and two-state models of gene expression showed both the potential and limitations of these models to interpret stochastic gene regulation. The burst model described the data well under conditions of low expression, when the gene was active for ≤5% of the mean protein lifetime, but diverged for increasing expression levels. The two-state model better described the data at higher levels of expression, but near full induction the error in the activation and inactivation rates became significant. Additionally, the fits provided estimates of the number of cell measurements necessary to produce reliable parameter estimates. With 50 cells the worst-case relative error was ±90%, but with 200 cells it dropped to ±32%. Fitting to joint mRNA–protein distributions might improve parameter estimation. Fits to data with positive feedback indicated that both models were unable to reliably extract parameters from populations with such feedback.

Switching of cells from the uninduced to the induced state was observed in the positive feedback model without DNA looping over a range of low inducer concentrations. During switching, the uninduced population maintained a stable stationary distribution while cells stochastically transitioned to the induced population. The probability landscape showed that both an mRNA and a protein threshold must be crossed for a cell to switch to the induced state. The probability landscape for the DNA looping case is likely different, but additional model states would be required to accurately represent DNA looping.

Finally, we have presented what we believe to be the first whole-cell simulations of stochastic gene expression using experimentally obtained cellular architecture. These simulations showed that in vivo conditions can impact the stochastic noise in biological systems. Positional memory of transcription factors following unbinding, amplified by anomalous diffusion due to molecular crowding, introduces non-Poissonian statistics. In the case of our lac model in fast-growth cells, this effect caused a decrease in the mean value of the LacY distribution by up to 10% and its noise by up to 20%, for a given environmental condition. In a slow-growth cell phenotype we saw a large increase in burst frequency due to the smaller cell size, as determined from cryoelectron tomography. From this difference we infer that changes in cellular size and/or shape during the cell cycle can have an impact on stochastic processes. Since spatial noise can vary from cell-to-cell or even during the cell cycle so we consider it a type of extrinsic noise. The necessary computational resources and experimental data are becoming available such that computational biologists should consider adding spatial degrees of freedom into physical models of cellular biochemical networks.

Supporting Information

Text S1

Supporting methods. Further methods describing the derivation of the rate constant relationship for non-cooperative ligand binding, the lattice microbe reaction operator, and the lattice coarse-graining technique.

(PDF)

Video S1

Simulated colony of E. coli cells responding to inducer. Video composite of trajectories from six spatial PFB+IV simulations at 15 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e421.jpg inducer. Yellow circles are LacY proteins and red circle are mY mRNA molecules. Two cells begin the process of switching to the induced state.

(MOV)

Video S2

Trajectory of a PFB+IV+CET cell responding to inducer. Visualization of a single slow-growth CET modeled cell responding to 15 An external file that holds a picture, illustration, etc.
Object name is pcbi.1002010.e422.jpg inducer. Gray spheres are ribosomes and the blue region the nucleoid. Yellow circles are LacY proteins and red circles are mY mRNA molecules. The repressor–operator complex is green and the free operator is white.

(MOV)

Acknowledgments

The authors would like to acknowledge John Stone for helpful discussions on implementation of the code on the GPU.

Footnotes

The authors have declared that no competing interests exist.

This research was supported by the Department of Energy Office of Science (BER), the National Science Foundation (MCB-0844670, PHY-0822613), and the Foundation Fourmentin-Guilbert. Computational resources were provided by the NSF through the TeraGrid and NCSA (TG-MCA03S027) and also by the CUDA Center of Excellence at UIUC. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

References

1. Thattai M, van Oudenaarden A. Intrinsic noise in gene regulatory networks. Proc Natl Acad Sci USA. 2001;98:8614–9. [PMC free article] [PubMed]
2. Elowitz MB, Levine AJ, Siggia ED, Swain PS. Stochastic gene expression in a single cell. Science. 2002;297:1183–6. [PubMed]
3. Swain PS, Elowitz MB, Siggia ED. Intrinsic and extrinsic contributions to stochasticity in gene expression. Proc Natl Acad Sci USA. 2002;99:12795–800. [PMC free article] [PubMed]
4. Paulsson J. Summing up the noise in gene networks. Nature. 2004;427:415–8. [PubMed]
5. Thattai M, van Oudenaarden A. Stochastic gene expression in fluctuating environments. Genetics. 2004;167:523–30. [PMC free article] [PubMed]
6. Acar M, Mettetal JT, van Oudenaarden A. Stochastic switching as a survival strategy in fluctuating environments. Nat Genet. 2008;40:471–5. [PubMed]
7. Paulsson J. Models of stochastic gene expression. Phys Life Rev. 2005;2:157–175.
8. Kaern M, Elston TC, Blake WJ, Collins JJ. Stochasticity in gene expression: from theories to phenotypes. Nat Rev Genet. 2005;6:451–64. [PubMed]
9. Raj A, van Oudenaarden A. Nature, nurture, or chance: stochastic gene expression and its consequences. Cell. 2008;135:216–26. [PMC free article] [PubMed]
10. Shahrezaei V, Swain P. The stochastic nature of biochemical networks. Curr Opin Biotechnol. 2008;19:369–374. [PubMed]
11. Cheong R, Paliwal S, Levchenko A. Models at the single cell level. Wiley Interdiscip Rev Syst Biol Med. 2009;2:34–48. [PMC free article] [PubMed]
12. McQuarrie D. Stochastic approach to chemical kinetics. J Appl Probab. 1967;4:413–478.
13. Paulsson J, Berg OG, Ehrenberg M. Stochastic focusing: fluctuation-enhanced sensitivity of intracellular regulation. Proc Natl Acad Sci USA. 2000;97:7148–53. [PMC free article] [PubMed]
14. Hasty J, Pradines J, Dolnik M, Collins JJ. Noise-based switches and amplifiers for gene expression. Proc Natl Acad Sci USA. 2000;97:2075–80. [PMC free article] [PubMed]
15. Sasai M, Wolynes PG. Stochastic gene expression as a many-body problem. Proc Natl Acad Sci USA. 2003;100:2374–9. [PMC free article] [PubMed]
16. Gillespie DT. Stochastic simulation of chemical kinetics. Annu Rev Phys Chem. 2007;58:35–55. [PubMed]
17. McAdams HH, Arkin A. Stochastic mechanisms in gene expression. Proc Natl Acad Sci USA. 1997;94:814–9. [PMC free article] [PubMed]
18. Arkin A, Ross J, McAdams HH. Stochastic kinetic analysis of developmental pathway bifurcation in phage lambda-infected Escherichia coli cells. Genetics. 1998;149:1633–48. [PMC free article] [PubMed]
19. Ozbudak EM, Thattai M, Kurtser I, Grossman AD, van Oudenaarden A. Regulation of noise in the expression of a single gene. Nat Genet. 2002;31:69–73. [PubMed]
20. Blake WJ, Kaern M, Cantor CR, Collins JJ. Noise in eukaryotic gene expression. Nature. 2003;422:633–7. [PubMed]
21. Yu J, Xiao J, Ren X, Lao K, Xie XS. Probing gene expression in live cells, one protein molecule at a time. Science. 2006;311:1600–3. [PubMed]
22. Choi PJ, Cai L, Frieda K, Xie XS. A stochastic single-molecule event triggers phenotype switching of a bacterial cell. Science. 2008;322:442–6. [PMC free article] [PubMed]
23. Raj A, Rifkin SA, Andersen E, van Oudenaarden A. Variability in gene expression underlies incomplete penetrance. Nature. 2010;463:913–918. [PMC free article] [PubMed]
24. Friedman N, Cai L, Xie XS. Linking stochastic dynamics to population distribution: an analytical framework of gene expression. Phys Rev Lett. 2006;97:168302. [PubMed]
25. Shahrezaei V, Swain PS. Analytical distributions for stochastic gene expression. Proc Natl Acad Sci USA. 2008;105:17256–61. [PMC free article] [PubMed]
26. Choi PJ, Xie XS, Shakhnovich EI. Stochastic switching in gene networks can occur by a single-molecule event or many molecular steps. J Mol Biol. 2010;396:230–44. [PubMed]
27. van Zon JS, Morelli MJ, Tănase-Nicola S, ten Wolde PR. Diffusion of transcription factors can drastically enhance the noise in gene expression. Biophys J. 2006;91:4350–67. [PMC free article] [PubMed]
28. Morelli MJ, tenWolde PR. Reaction Brownian dynamics and the effect of spatial fluctuations on the gain of a push-pull network. J Chem Phys. 2008;129:054112. [PubMed]
29. Takahashi K, Tanase-Nicola S, ten Wolde PR. Spatio-temporal correlations can drastically change the response of a MAPK pathway. Proc Natl Acad Sci USA. 2010;107:2473–8. [PMC free article] [PubMed]
30. Ortiz JO, Förster F, Kürner J, Linaroudis AA, Baumeister W. Mapping 70S ribosomes in intact cells by cryoelectron tomography and pattern recognition. J Struct Biol. 2006;156:334–41. [PubMed]
31. Comolli LR, Baker BJ, Downing KH, Siegerist CE, Banfield JF. Three-dimensional analysis of the structure and ecology of a novel, ultra-small archaeon. ISME J. 2009;3:159–67. [PubMed]
32. Briegel A, Ortega DR, Tocheva EI, Wuichet K, Li Z, et al. Universal architecture of bacterial chemoreceptor arrays. Proc Natl Acad Sci USA. 2009;106:17181–6. [PMC free article] [PubMed]
33. Beck M, Malmström JA, Lange V, Schmidt A, Deutsch EW, et al. Visual proteomics of the human pathogen Leptospira interrogans. Nat Meth. 2009;6:817–23. [PMC free article] [PubMed]
34. Kühner S, van Noort V, Betts MJ, Leo-Macias A, Batisse C, et al. Proteome organization in a genome-reduced bacterium. Science. 2009;326:1235–40. [PubMed]
35. Banks DS, Fradin C. Anomalous diffusion of proteins due to molecular crowding. Biophys J. 2005;89:2960–71. [PMC free article] [PubMed]
36. Golding I, Cox EC. Physical nature of bacterial cytoplasm. Phys Rev Lett. 2006;96:098102. [PubMed]
37. Roberts E, Stone JE, Sepulveda L, Hwu WMW, Luthey-Schulten Z. Proceedings of the 2009 IEEE International Symposium on Parallel & Distributed Processing. Los Alamitos, CA, USA: IEEE Computer Society; 2009. Long time-scale simulations of in vivo diffusion using GPU hardware. pp. 1–8.
38. Ridgway D, Broderick G, Lopez-Campistrous A, Ru'aini M, Winter P, et al. Coarse-grained molecular simulation of diffusion and reaction kinetics in a crowded virtual cytoplasm. Biophys J. 2008;94:3748–59. [PMC free article] [PubMed]
39. McGuffee SR, Elcock AH. Diffusion, crowding & protein stability in a dynamic molecular model of the bacterial cytoplasm. PLoS Comput Biol. 2010;6:e1000694. [PMC free article] [PubMed]
40. Andrews SS, Bray D. Stochastic simulation of chemical reactions with spatial resolution and single molecule detail. Phys Biol. 2004;1:137–51. [PubMed]
41. van Zon JS, ten Wolde PR. Green's-function reaction dynamics: a particle-based approach for simulating biochemical networks in time and space. J Chem Phys. 2005;123:234910. [PubMed]
42. Hattne J, Fange D, Elf J. Stochastic reaction-diffusion simulation with MesoRD. Bioinformatics. 2005;21:2923–4. [PubMed]
43. Rodríguez JV, Kaandorp JA, Dobrzy'nski M, Blom JG. Spatial stochastic modelling of the phosphoenolpyruvate-dependent phosphotransferase (PTS) pathway in Escherichia coli. Bioinformatics. 2006;22:1895–901. [PubMed]
44. Arjunan SNV, Tomita M. A new multicompartmental reaction-diffusion modeling method links transient membrane attachment of E coli MinE to E-ring formation. Syst Synth Biol. 2010;4:35–53. [PMC free article] [PubMed]
45. Golding I, Paulsson J, Zawilski SM, Cox EC. Real-time kinetics of gene activity in individual bacteria. Cell. 2005;123:1025–36. [PubMed]
46. Yildiz A, Forkey JN, McKinney SA, Ha T, Goldman YE, et al. Myosin V walks hand-overhand: single fluorophore imaging with 15-nm localization. Science. 2003;300:2061–5. [PubMed]
47. Betzig E, Patterson GH, Sougrat R, Lindwasser OW, Olenych S, et al. Imaging intracellular fluorescent proteins at nanometer resolution. Science. 2006;313:1642–5. [PubMed]
48. Huang B, Wang W, Bates M, Zhuang X. Three-dimensional super-resolution imaging by stochastic optical reconstruction microscopy. Science. 2008;319:810–3. [PMC free article] [PubMed]
49. Biteen JS, Thompson MA, Tselentis NK, Bowman GR, Shapiro L, et al. Super-resolution imaging in live Caulobacter crescentus cells using photoswitchable EYFP. Nat Meth. 2008;5:947–9. [PMC free article] [PubMed]
50. Ortiz JO, Brandt F, Matias VRF, Sennels L, Rappsilber J, et al. Structure of hibernating ribosomes studied by cryoelectron tomography in vitro and in situ. J Cell Biol. 2010;190:613–21. [PMC free article] [PubMed]
51. Elf J, Li GW, Xie XS. Probing transcription factor dynamics at the single-molecule level in a living cell. Science. 2007;316:1191–4. [PMC free article] [PubMed]
52. Wong P, Gladney S, Keasling JD. Mathematical model of the lac operon: inducer exclusion, catabolite repression, and diauxic growth on glucose and lactose. Biotechnol Progr. 1997;13:132–43. [PubMed]
53. Vilar JMG, Guet CC, Leibler S. Modeling network dynamics: the lac operon, a case study. J Cell Biol. 2003;161:471–6. [PMC free article] [PubMed]
54. van Hoek M, Hogeweg P. The effect of stochasticity on the lac operon: an evolutionary perspective. PLoS Comput Biol. 2007;3:e111. [PMC free article] [PubMed]
55. Kuhlman T, Zhang Z, Saier MH, Hwa T. Combinatorial transcriptional control of the lactose operon of Escherichia coli. Proc Natl Acad Sci USA. 2007;104:6043–8. [PMC free article] [PubMed]
56. Narang A. Effect of DNA looping on the induction kinetics of the lac operon. J Theor Biol. 2007;247:695–712. [PubMed]
57. Santillán M. Bistable behavior in a model of the lac operon in Escherichia coli with variable growth rate. Biophys J. 2008;94:2065–81. [PMC free article] [PubMed]
58. Noel JT, Pilyugin SS, Narang A. The diffusive influx and carrier efflux have a strong effect on the bistability of the lac operon in Escherichia coli. J Theor Biol. 2009;256:14–28. [PubMed]
59. Bell CE, Lewis M. A closer view of the conformation of the Lac repressor bound to operator. Nat Struct Biol. 2000;7:209–14. [PubMed]
60. Lewis M. The lac repressor. C R Biol. 2005;328:521–48. [PubMed]
61. Xu J, Matthews KS. Flexibility in the inducer binding region is crucial for allostery in the Escherichia coli lactose repressor. Biochemistry. 2009;48:4988–98. [PMC free article] [PubMed]
62. Mettetal JT, Muzzey D, Pedraza JM, Ozbudak EM, van Oudenaarden A. Predicting stochastic gene expression dynamics in single cells. Proc Natl Acad Sci USA. 2006;103:7304–9. [PMC free article] [PubMed]
63. Stamatakis M, Mantzaris NV. Comparison of deterministic and stochastic models of the lac operon genetic network. Biophys J. 2009;96:887–906. [PMC free article] [PubMed]
64. Oehler S, Alberti S, Müller-Hill B. Induction of the lac promoter in the absence of DNA loops and the stoichiometry of induction. Nucleic Acids Res. 2006;34:606–12. [PMC free article] [PubMed]
65. Condon C. Maturation and degradation of RNA in bacteria. Curr Opin Microbiol. 2007;10:271–8. [PubMed]
66. Akiyama Y. Quality control of cytoplasmic membrane proteins in Escherichia coli. J Biochem. 2009;146:449–54. [PubMed]
67. Royer CA, Chakerian AE, Matthews KS. Macromolecular binding equilibria in the lac repressor system: studies using high-pressure fluorescence spectroscopy. Biochemistry. 1990;29:4959–66. [PubMed]
68. Oehler S, Eismann ER, Krämer H, Müller-Hill B. The three operators of the lac operon cooperate in repression. EMBO J. 1990;9:973–9. [PMC free article] [PubMed]
69. Maloney PC, Wilson TH. Quantitative aspects of active transport by the lactose transport system of Escherichia coli. Biochim Biophys Acta. 1973;330:196–205. [PubMed]
70. Dornmair K, Overath P, Jähnig F. Fast measurement of galactoside transport by lactose permease. J Biol Chem. 1989;264:342–6. [PubMed]
71. Kepes A. The B-galactoside permease of Escherichia coli. J Membrane Biol. 1971;4:87–112. [PubMed]
72. Kepes A. [Kinetic studies on galactoside permease of Escherichia coli]. Biochim Biophys Acta. 1960;40:70–84. [PubMed]
73. Ohshima Y, Mizokoshi T, Horiuchi T. Binding of an inducer to the lac repressor. J Mol Biol. 1974;89:127–36. [PubMed]
74. O’Gorman RB, Rosenberg JM, Kallai OB, Dickerson RE, Itakura K, et al. Equilibrium binding of inducer to lac repressoroperator DNA complex. J Biol Chem. 1980;255:10107–14. [PubMed]
75. Dunaway M, Olson JS, Rosenberg JM, Kallai OB, Dickerson RE, et al. Kinetic studies of inducer binding to lac repressoroperator complex. J Biol Chem. 1980;255:10115–9. [PubMed]
76. Barkley MD, Riggs AD, Jobe A, Burgeois S. Interaction of effecting ligands with lac repressor and repressor-operator complex. Biochemistry. 1975;14:1700–12. [PubMed]
77. Karapiperis T, Blankleider B. Cellular automaton model of reaction-transport processes. Phys D. 1994;78:30–64.
78. Driessen AJM, Nouwen N. Protein translocation across the bacterial cytoplasmic membrane. Annu Rev Biochem. 2008;77:643–67. [PubMed]
79. Macfarlane J, Müller M. The functional integration of a polytopic membrane protein of Escherichia coli is dependent on the bacterial signal-recognition particle. Eur J Biochem. 1995;233:766–71. [PubMed]
80. Seluanov A, Bibi E. FtsY, the prokaryotic signal recognition particle receptor homologue, is essential for biogenesis of membrane proteins. J Biol Chem. 1997;272:2053–5. [PubMed]
81. Facey SJ, Kuhn A. Membrane integration of E coli model membrane proteins. Biochim Biophys Acta. 2004;1694:55–66. [PubMed]
82. Fusco D, Accornero N, Lavoie B, Shenoy SM, Blanchard JM, et al. Single mRNA molecules demonstrate probabilistic movement in living mammalian cells. Curr Biol. 2003;13:161–7. [PubMed]
83. Ishihama Y, Funatsu T. Single molecule tracking of quantum dot-labeled mRNAs in a cell nucleus. Biochem Biophys Res Comm. 2009;381:33–8. [PubMed]
84. Gowrishankar J, Harinarayanan R. Why is transcription coupled to translation in bacteria? Mol Microbiol. 2004;54:598–603. [PubMed]
85. Taniguchi Y, Choi PJ, Li GW, Chen H, Babu M, et al. Quantifying E coli Proteome and Transcriptome with Single-Molecule Sensitivity in Single Cells. Science. 2010;329:533–8. [PMC free article] [PubMed]
86. Setty Y, Mayo AE, Surette MG, Alon U. Detailed map of a cis-regulatory input function. Proc Natl Acad Sci USA. 2003;100:7702–7. [PMC free article] [PubMed]
87. Rosenfeld N, Young JW, Alon U, Swain PS, Elowitz MB. Gene regulation at the single-cell level. Science. 2005;307:1962–5. [PubMed]
88. Mayo AE, Setty Y, Shavit S, Zaslaver A, Alon U. Plasticity of the cis-regulatory input function of a gene. PLoS Biol. 2006;4:e45. [PMC free article] [PubMed]
89. Woldringh CL, de Jong MA, van den Berg W, Koppes L. Morphological analysis of the division cycle of two Escherichia coli substrains during slow growth. J Bacteriol. 1977;131:270–9. [PMC free article] [PubMed]
90. Llopis PM, Jackson AF, Sliusarenko O, Surovtsev I, Heinritz J, et al. Spatial organization of the flow of genetic information in bacteria. Nature. 2010;466:77–81. [PMC free article] [PubMed]
91. Raj A, Peskin CS, Tranchina D, Vargas DY, Tyagi S. Stochastic mRNA synthesis in mammalian cells. PLoS Biol. 2006;4:e309. [PMC free article] [PubMed]
92. Goeddel DV, Yansura DG, Caruthers MH. Binding of synthetic lactose operator DNAs to lactose represessors. Proc Natl Acad Sci USA. 1977;74:3292–6. [PMC free article] [PubMed]
93. Chung J, Stephanopoulos G. On physiological multiplicity and population heterogeneity of biological systems. Chem Eng Sci. 1996;51:1509–1521.

Articles from PLoS Computational Biology are provided here courtesy of Public Library of Science
PubReader format: click here to try

Formats:

Related citations in PubMed

See reviews...See all...

Cited by other articles in PMC

See all...

Links