Synthetic approaches to study transcriptional networks and noise in mammalian systems

Abstract

Synthetic biology aims to build new functional organisms and to rationally re‐design existing ones by applying the engineering principle of modularity. Apart from building new life forms to perform technical applications, the approach of synthetic biology is useful to dissect complex biological phenomena into simple and easy to understand synthetic modules. Synthetic gene networks have been successfully implemented in prokaryotes and lower eukaryotes, with recent approaches moving ahead towards the mammalian environment. However, synthetic circuits in higher eukaryotes present a more challenging scenario, since its reliability is compromised because of the strong stochastic nature of transcription. Here, the authors review recent approaches that take advantage of the noisy response of synthetic regulatory circuits to learn key features of the complex machinery that orchestrates transcription in higher eukaryotes. Understanding the causes and consequences of biological noise will allow us to design more reliable mammalian synthetic circuits with revolutionary medical applications.

1 Introduction

Synthetic biology aspires to engineer novel biological circuits that exhibit new programmed behaviour taking advantage of well‐characterised biomolecular components and genetic modules [1, 2]. The modular approach of synthetic biology proposes the use of small biological elements as single interchangeable components that are reconnected to form larger systems [3]. Ideally, such a system could be decomposed into its constituent building blocks to be recombined later in order to develop new functions [4, 5, 6].

Several of the pioneering works in the field of synthetic biology focused on understanding the basis of gene regulation by implementing exogenous gene circuits in bacteria, such as the bistable toggle switch [7] or the oscillatory repressilator [8]. Nonetheless, the rational assembly of these simplified genetic networks in higher eukaryotes represents a more challenging task, since their complex regulatory machinery induces very strong fluctuations and a highly unpredictable output in terms of protein levels, dynamics and cell‐to‐cell variability [9].

Whereas synthetic biology has limited success in mammalian cells so far, the simplicity of synthetic network modules can be used to study key aspects in mammalian transcription. This review focuses on several studies that benefit from simple synthetic designs to understand basic principles of mammalian regulation of transcription. The first section deals with a synthetic linear network, providing strong experimental evidences of mammalian transcriptional noise. The second case explores the noisy signal output of a non‐linear oscillatory network, and tries to supply further mathematical reasoning for its stochastic behaviour. Finally, the last section explores the interplay between noise in transcription and autoregulatory motifs common in biological networks. These few selected examples illustrate how the powerful combination of synthetic networks with simple mathematical models offers an advantageous workflow to unveil the causes and consequences of noise and variability in mammalian systems.

2 Synthetic construct to understand transcriptional bursting

It is well documented that noise in gene expression induces variability in genetically identical cell populations, even under the same external conditions [9]. Simple synthetic gene regulatory networks have been utilised to investigate this phenomenon in model organisms such as Escherichia coli [10] and Saccharomyces cerevisiae [11]. To investigate the characteristics of noise in higher eukaryotes, Raj et al. [12] designed a synthetic circuit to study the variation of mRNA and protein production within clonal populations. They implemented a synthetic reporter gene with multiple copies of a probe‐binding sequence at the 3′ end of the coding sequence for a fluorescent protein (Fig. 1). The gene was cloned downstream of a minimal promoter regulated by the doxycycline sensitive tet operator. Finally, the construct was integrated into Chinese Hamster Ovary (CHO) cells, allowing the researchers to monitor individual mRNA molecules at different transcription rates using fluorescence in situ hybridisation (FISH).

Fig. 1.

Synthetic network to study transcriptional bursting in mammalian cells

a Diagram of the synthetic construct developed by Raj et al. [12]. The cell line CHO was engineered to express the tetracycline‐controlled Tet‐off transactivator tTA (a fusion of the E. coli TetR repressor and the Herpes simplex VP16) which binds to the tet operator sequence and promotes the transcription of the YFP. The antibiotic doxycycline, which prevents tTA to bind to the operator sequence, is added to the growth medium to regulate the level of produced mRNA. The probe binding sequence detected by fluorescent in situ hybridisation FISH provides single molecule sensitivity

b Typical output of transcriptional burst in the mRNA production

Experiments revealed that, under constant conditions, mRNA production occurs at a variable rate during sudden brief periods of time in a burst‐like process. Statistical analysis of the experimental data showed that variability (defined as the standard deviation of the production pulse normalised by the mean) remained constant for cultures under different doxycycline levels, meaning that noise was independent on the mRNA concentration. This result contradicts the basic property predicted by stochastic models for transcription [13], where the deviation from the mean should decrease when the number of interacting molecules increases, and the system should tend to its deterministic limit [14].

To further investigate this matter, several constructs were engineered with different number of copies of the tet operator sequence. This allowed them to test if mRNA production increases proportionally with the number of binding sites. Results showed that mRNA levels correlated with the number of active transcription sites, but only affecting the burst size (i.e. the number of mRNA produced in a transcription event) and not the frequency of bursting (i.e. the chances of being transcribed). This evidenced again that other intrinsic random mechanism different than fluctuations is inducing stochasticity in gene activation. Therefore conventional stochastic models used to explain intrinsic noise in bacterial gene expression are not applicable to the study of mammalian transcription.

Based on the experimental observations, the authors proposed a mathematical model that includes random activation/inactivation of transcription [15, 16, 17]. The model consisted of a gene which transitions randomly between an active state A with high rate of production of mRNA molecules (denoted M), and an inactive state I, in which transcription is not possible. The time during which the gene is in the active state A is the burst of mRNA synthesis. In this model, the following reactions are considered

$$I\stackrel{\lambda }{\to }A$$

$$A\stackrel{\mu }{\to }I$$

$$A\stackrel{\gamma }{\to }A+M$$

$$M\stackrel{\delta }{\to }Ø$$

where λ is the rate of gene activation, γ is the rate of gene inactivation, μ is the rate of transcription when the gene is in the active state, and δ is the rate of mRNA decay. These reactions generate the stochastic chemical master equation that describes the variation in time of the probability of having n molecules of mRNA at a given time. Fitting theoretical parameters to the experimental distributions allows the measurement of the rate of gene activation and the average burst size.

An important conclusion provided by the experimental data in combination with the mathematical model is that another source of stochasticity, besides random chemical reactions typical of bacterial noise, operates in mammalian transcription. Although several other studies have encounter this feature in higher eukaryotes [18, 19], little is known about the possible causes of bursting. According to Raj et al., one potential explanation for transcriptional bursts may reside in chromosomal positioning [20]. If random events of chromatin remodelling were proven as a feasible burst‐generating mechanism, then, the activation of a gene would be facilitated by chromatin de‐condensation, whereas gene inactivation would correspond to chromatin re‐condensation. This hypothesis is in all ways consistent with experimental and theoretical observations, still, some other compatible mechanisms may also explain the emergence of bursts, as discussed below.

3 Synthetic network to study polymerase trafficking

Transcription is tightly regulated by a combination of elements such as promoter architecture, number of transcription factor binding sites, chromatin structure, amount of DNA‐binding proteins and co‐factors of the transcriptional complex [21]. Intronic non‐coding DNA is itself a core regulator of transcription where its main function is to separate coding regions for genes to be processed appropriately. In addition, introns are also thought to introduce important time delays in mRNA synthesis and, consequently, in protein production [22]. This temporal regulation is known to be critical, for instance, during early development where precise timing is a basic demand [23].

Swinburne et al. [24] developed a synthetic network to study the effect of intronic DNA in the dynamics of gene regulation. They implemented a synthetic circuit consisting of a fluorescent protein engineered to repress its own transcription. As depicted in Fig. 2 a, the yellow fluorescent protein (YFP) fused to the Tet repressor (TetR) is expressed under the control of a minimal promoter regulated by the tet‐O sequence. Several cell lines were established with different lengths of the intronic DNA between the promoter and the YFP gene. This way, Swinburne et al. created a network where the gene inhibition is performed by the gene output with a delay set by the intron length [25, 26].

Fig. 2.

Mammalian synthetic negative feedback loop with variable delay

a Diagram of the synthetic construct developed by Swinburne et al. [24]. An oscillatory behaviour is achieved as a result of an imposed delay in an autoinhibitory system. TetR binds to the tet‐O sequence and inhibits the transcription of the YFP that is used to monitor the network dynamics. A couple of exogenous intronic sequences are inserted after the first intron of the promoter (1 kb). The modular intron sequences introduced were either 7 or 16 kb long, providing three different types of cells with the same coding region

b Mathematical modelling of the synthetic network include explicit simulations of the RNA polymerase II transcription elongation, that was forced to move along vector spaces equivalent to either 3, 10 or 19 kb. Simulated polymerases transcribed at constant velocities sampled from a Gaussian distribution with a mean velocity of 1 kb/min and a σ = 0.255 kb/min. First column represents protein output for different gene length. Oscillations varied with gene length showing that transcriptional bursting is promoted by long polymers chains. As shown in the centre and right columns, time between pulses and protein levels, respectively, depicted a broader distributions of values when increasing gene length, proving that congestion events produce large heterogeneities in the system dynamics

On the other hand, it is well‐known that under the right conditions, a delayed negative feedback interaction results in oscillations in protein and mRNA levels. In this case, Swinburne et al. obtained a noisy response with large cell‐to‐cell variation, and most importantly with single‐cell oscillations occurring in a burst‐like manner. The period of the oscillations varied from cell‐to‐cell, even in clonal populations, but the mean values correlated well with the length of the inserted DNA (i.e. longer genes produced longer periods). Interestingly, the experimental results showed that the length of the intron correlated with the standard deviation of the oscillatory period as well, meaning that some mechanism involved in transcription is responsible for the noisy response in this system. This experimental result cannot be explained by the hypotheses of bursts generated by random chromatin de‐condensation. Chromatin variations do not account for the changes in the standard deviation of the period, since they would affect equally all genes independently of their length.

Swinburne et al. developed a phenomenological model including the experimental fact that polymerases travel throughout the DNA at different velocities [27, 28]. The existence of slow and fast transcription elongation rates eventually results in a stack of polymerases piling up behind a slow polymerase, with the transcriptional bursting caused by a ‘traffic jam’ phenomenon [29]. In consequence, the stack of polymerases terminating transcription elongation will result in a sudden peak in production of mRNA.

In this case, the biological system is described by means of a deterministic mathematical model with ordinary differential equations, following a mass balance law that defines the rate of change of the concentration of a given species S over time as the rate of reactions where the molecule is produced minus the rate of the reactions where it is degraded. The following differential equations describe the trajectories of the protein, p, and mRNA, m

$$\frac{\mathrm{d}p(t)}{\mathrm{d}t}=am[t-T_p]-bp(t)$$

$$\frac{\mathrm{d}m(t)}{\mathrm{d}t}=\frac{n_{\text{pol}}}{1+\left(p^2[t-T_m]/\left(p_0^2\right)\right)}-cm(t)$$

Parameter a represents the rate of protein production, b and c are degradation rates (inverse lifetimes), T_p and T_m account for the delay related to translation and transcription, respectively, and n _pol is the number of polymerases that terminate elongation at a given time, modelled as traveling through the gene with different velocities [24].

Numerical simulations were performed for the different gene lengths used in the experiments (3, 10, 19 kb) considering the experimental measured Gaussian velocity distribution for the polymerases [30], showing that bursting becomes more pronounced as DNA length increases, in agreement with the experimental results. Although the chance of producing a slow‐leading polymerase is the same regardless of gene length, the fact that polymerases have to travel throughout longer introns increases the probability over time of encountering a dilatory polymerase.

To gain further insight into the mechanism and the effect of polymerase dynamic heterogeneities as a burst‐generating mechanism, we used the model developed in [24] to compute additional numerical simulations. Fig. 2 b shows single trajectories of the protein for increasing values of gene length. The distribution values for the time between peaks broadens with the intron length, as well as the distribution of the amplitude of the peaks. The effect of the traffic jam reduces the mean value of the polymerase velocity, increasing the mean value of the period whereas inducing peaks of very high amplitude (2 orders of magnitude increase).

As depicted in Fig. 3, the typical sinusoidal oscillatory behaviour is retrieved when homogeneous velocity among the polymerases is considered. Increasing the standard deviation, induces large pulse‐to‐pulse variability in protein production, leading to a difference of two orders of magnitude in the amplitude of the peak compared to the amplitude in the purely oscillatory regime. Fig. 3 b shows how the variability in polymerase velocity induces pauses in the rate of polymerases terminating transcription, followed by sudden arrival of many polymerases at once, in contrast with the constant rate in the purely oscillatory regime (Fig. 3 a).

Fig. 3.

Simulations of polymerase traffic jams with different velocity distributions for a 10 kb gene

a For standard deviation σ = 0, polymerases transcribe and terminate at a constant rate leading to regular oscillations

b When σ = 0.44 kb/min, the assignation of random broad velocities produces irregular pulses and long pauses. The slow polymerase object creates a congestion where the pause is correlated with the amount of polymerases terminating at the same time

Altogether, the experimental and modelling results revealed that both gene length and velocity distribution may be responsible for the bursty behaviour. Other post‐transcriptional processes might also be influencing the retardation times by altering, for instance, the splicing rates. However, they were not included in the simulation since the dynamics of splicing is much faster than transcription, and this process is theoretically gene‐length invariant [31].

4 Synthetic autoregulatory networks and noise

The two previous cases evidence the importance of transcriptional intrinsic noise in mammalian systems. The experiments reviewed also suggest that the source of stochasticity in higher eukaryotes may be different from the one in bacteria, normally associated with the low numbers of interacting molecules [32]. Whereas the first example focuses on noise arising from a synthetic linear network, the second example deals with its propagation in a syntetic nonlinear network that exhibits oscillations. Since these nonlinear autoregulatory motifs are ubiquitous in prokaryotes and eukaryotes [33], it is key to understand how these feedback systems interplay with the intrinsic stochastic characteristics of transcriptional noise [34, 35, 36].

One of the simplest autoregulatory motifs is a positive feedback loop, formed by a protein that promotes its own production. A positive feedback generally turns the graded response of linear architectures into a bistable expression pattern with two distinct solutions for the same parameter values: an ‘on’ state where the gene expression is maximal and an ‘off’ state where transcription occurs at a low rate [37]. Systems that present bistability are also hysteretic, meaning that the required signal strength to switch from one state to another depends on the initial state. Therefore switching from the ‘off’ to the ‘on’ state in a bistable system requires a higher inducer concentration than the one required to switch from ‘on’ to ‘off’ [38, 39].

A neat example of the synthetic implementation of a positive feedback loop was developed by Kramer et al. [40] (Fig. 4), where they use mathematical analysis to understand how cooperativity in the tTA (formed by a fusion of the E. coli TetR repressor and the Herpes simplex VP16) is a requirement for hysteresis. Other cases oriented to clinical purposes include the experiment conducted by May et al., who used a positive feedback loop to control cell proliferation, as a potential tool for tissue engineering in human cells [41]. In addition, Burril et al. designed as well as a synthetic autocatalytic network that resulted in a memory device used to track human cells which responded to a specific stimulus linked to endogenous hypoxia and DNA damage response pathways [42].

Fig. 4.

Synthetic positive feedback loop leading to a bimodal and hysteretic expression pattern

a Scheme of the circuit implemented by Kramer et al. in CHO cells [40]. The tetracycline dependent transactivator (tTA) induces the hybrid promoter tet ₀₇ − ETR₈ − P _hCMVmin driving its own expression as well as the expression of SEAP (human placental secreted alkaline phosphatase), used to quantify the expression of tTA. The constitutively expressed transrepressor E‐KRAB (a fusion of the E. coli macrolide resistance operon repressor E and the human trans‐silencing domain KRAB) binds to the ETR₈ region of the promoter in a fashion modulated by the antibiotic erythromycin (EM) who inhibits the transrepressor capacity by binding also to the hybrid promoter. In the absence of antibiotics the E‐KRAB action dominates over the tTA function resulting in the suppression of the positive feedback loop

b Profile of the gene expression provided by the synthetic arrangement. For the same stimulus two possible responses are possible (bistability) and the signal intensity to switch from one state to another depends on the initial state (hysteresis)

These contributions do not explicitly focus on the consequences of noisy transcription in positively autoregulated systems, but the findings in other model organisms suggest interesting interplay between feedback and noise. In bacteria, synthetic positive feedbacks induce an increase in the sensitivity to perturbations, leading to high cell‐to‐cell variability in clonal populations [43]. Although when noise levels are low, bistable systems have the ability to buffer against modest changes, preventing the system to randomly switch between states. This can further be controlled synthetically by increasing the strength of the positive feedback loop, thus, producing constructs more resistant to small fluctuations [44]. However, significantly large fluctuations as depicted in mammalian cells could potentially promote switching between the two stable states of the system resulting in large variability [45].

The other type of basic autoregulatory motif is a negative feedback loop, constituted basically by a protein that directly or indirectly inhibits its own activity [46, 47]. Engineered self‐repressing networks have been implemented in prokaryotes such as E. coli [48, 49], but there has been no detailed study of the implications of synthetic negative feedback loops in a mammalian environment (the oscillatory network reviewed in the previous section made use of a negative feedback loop but did not focused explicitly on the implications of the nonlinear loop in the noisy signal). Experiments in E. coli showed an actual noise reduction in the output of such motifs when compared to linear networks, but only for certain values of the negative feedback strength. When repression is strong enough, the behaviour of the linear arrangement is retrieved [50]. It was also observed that negative autoregulation increases the network dynamics by speeding up the response to a perturbation [51]. Again, transferring this debate to a mammalian environment would be quite interesting, especially taking into account that the sources of stochasticity in higher eukaryotes may differ from those of lower organism. Besides, since negative feedback loops exhibit at some point enhanced robustness against fluctuations, they can be implemented to increase the reliability in the design of synthetic networks.

5 Discussion

Noise and stochasticity are intrinsic properties of mammalian biological systems [52]. However, it is still unclear if the sophisticated regulatory machinery in such systems has developed to benefit from noise or to overcome this problem [53]. It is possible that the actual answer is a combination of the two scenarios since it has been experimentally observed that noise is exploited in some cases [54, 55] and avoided in others [56, 57]. On a larger scale, the architecture of the biological network of interactions will be in charge of either smoothing or enhancing fluctuations in order to achieve a desired response.

Future perspectives for synthetic biology transcend mere isolated environments towards therapeutic applications for diverse medical purposes [58, 59]. In the last few years, many novel synthetic strategies emerged as pioneering examples of clinical applications, including bacteria engineered to target cancer cells [60, 61], or vaccines in the form of artificial vesicles containing synthetic DNA encoding for a functional antigen protein [62]. As the field progresses towards mammalian and multicellular organisms, efforts move towards the development of anticancer drugs [63], gene therapy [64] or tissue engineering [65].

Despite these and other achievements, mammalian synthetic biology has to overcome certain limitations before becoming a truly engineering discipline [66]. Firstly, the variability produced by the complex transcription/translation machinery in mammalian cells compromises the reliability of synthetic circuits [67]. As shown in the examples here reviewed, simple linear and oscillatory dynamics show highly nonlinear and non‐sinusoidal responses, respectively. As discussed here, the variability in the signal gets amplified or reduced by these autoregulated network motifs so, they could be useful to enhance the reliability of synthetic networks. The propagation of noise in nonlinear synthetic networks is itself a very interesting topic and needs to be understood in order to implement trustworthy genetic circuits in higher eukaryotes. In addition, the requirement of a standardised modularity is further restricted mainly because of the crosstalk of synthetic inserted pathways with natural existing ones [68]. To avoid this as much as possible, most synthetic pathways implemented in mammalian cells take advantage of bacterial antibiotic response regulators, such as the tetracycline‐controlled system from E. coli [69] that presents little interaction with the mammalian intercellular environment [70].

Although still far from the reliability of artificial networks in bacteria, synthetic biology has already revealed fundamental aspects of mammalian gene regulation. The inherent modularity of synthetic circuits facilitates their characterisation and analysis by relatively straightforward mathematical modelling, providing an optimal framework to study the basis of many complex biological processes.