The unique capabilities of biological organisms make them an attractive target for engineering. Living cells are self-replicating, self-repairing chemical factories that can, in principle, be reconfigured by altering their DNA blueprint. The development of DNA synthesis technology has allowed researchers to create any DNA sequence they desire, even entire genomes (Gibson et al. 2008). The ability to synthesize DNA promises to make the entire repertoire of known biological diversity available to synthetic biologists.

Despite this power, rationally designed biological devices rarely function entirely as predicted, and are often less robust than natural systems. With the price of DNA synthesis falling and worldwide interest in synthetic biology rising, the progress of synthetic biology has not been limited by funding or ambition. Instead, progress made in the last decade of synthetic biology research has revealed the unique difficulties of engineering living systems. In particular, endogenous regulatory systems often interfere with the function of synthetic biological devices (Arkin & Fletcher 2006).

Improving our ability to engineer biology will require an improved ability to predictably regulate biological systems. Natural biological systems are regulated at many levels: transcription; RNA processing; translation; protein–protein interactions; and protein–substrate interactions all exert control on cellular processes. The spatial and temporal organization of these myriad control systems is vital to their proper function. The fact that RNA interference, an integral control mechanism in eukaryotes, was only discovered in the late 1990s demonstrates that we are still discovering fundamental building blocks of biological systems (Fire et al. 1998).

Transcriptional control can be exploited for signal integration, i.e. processing multiple inputs and producing well-defined outputs. The operations that signal integration elements perform can be represented as logic gates. Boolean logic gates integrate two or more input signals and output one of two values (Irving 1961). For example, a two-input AND gate returns an output of ‘true’ when both inputs are true, and returns ‘false’ when either or both of the inputs are false (table 1). Transcriptional devices can serve as logic gates by producing outputs based on the state of input promoters; an AND gate would produce a certain output only if both input promoters were induced.

Ideally, signal integration elements and promoter inputs would be functionally separate in a synthetic device. For example, replacing an ara promoter with a lac promoter should modify the specificity of the device from arabinose to lactose without affecting how the rest of the device functions. Towards this end, a ‘modular’ AND gate was developed, allowing the modification of the promoter inputs and the gate output, leaving the core logic module intact (figure 1a; Anderson et al. 2007). Two promoters receive input to the device: one drives the expression of a T7 polymerase containing two amber mutations (T7ptag), while the other directs expression of the amber-suppressing tRNA supD. When both promoters are activated, supD expression allows the proper translation of T7ptag, which in turn transcribes a reporter gene with a T7 promoter. The original input promoters were induced by salicylate and arabinose for the expression of supD and T7ptag, respectively, with green fluorescent protein (GFP) as the output reporter.

This supD/T7ptag AND gate was intended to be modular, in that the device functions as an AND gate regardless of the input promoters or the output protein. To confirm modular function, the device inputs were replaced with magnesium-repressed and AI-1-inducible (LuxR quorum sensor) promoters, and the output was replaced with the invasin gene that triggers the invasion of mammalian cells. Although the initial reconfiguration of the device did not perform as an AND gate, tuning expression by varying the ribosome-binding sites of each promoter restored functionality. The new device allows Escherichia coli cells to invade mammalian cells in the case that exogenous magnesium is absent and the AI-1 signal is present. Importantly, the signal integration portion of the device was unmodified, and only minimal tuning was required to restore logic gate function after vastly altering the device inputs and outputs.

Producing high-fidelity outputs is as important as proper signal integration. Ideally, a synthetic device will remain in a persistent output state until the input state is changed. Genetic toggle switches have been demonstrated, incorporating transcriptional feedback loops to ensure that the device remains in one of two stable states (figure 1b; Gardner et al. 2000). Devices that feature oscillatory outputs based on transcriptional feedback have also been constructed (Elowitz & Leibler 2000; Stricker et al. 2008). ‘Memory devices’ feature a persistent output in response to a transient input; they are activated by a specific input and remain active after the input is removed (Ajo-Franklin et al. 2007).

Riboswitches are regulatory RNAs that bind small molecules or peptides via a specific aptamer domain (Patel et al. 1997). Natural riboswitches can be found in the 5′ untranslated region (UTR) of mRNA, where ligand binding triggers a conformational change that can negatively or positively modulate transcription or translation (Mandal & Breaker 2004; Tucker & Breaker 2005; Isaacs et al. 2006). Catalytic RNAs such as hammerhead ribozymes (HHRz; Khvorova et al. 2003) can even be attached to riboswitches to construct ligand-dependent ribozymes (Win & Smolke 2008).

In the context of synthetic devices, riboswitches allow promoter-independent control of gene expression. For example, RNA ‘antiswitches’ have been engineered in S. cerevisiae (Bayer & Smolke 2005). An antiswitch is an RNA molecule containing a ligand-binding aptamer domain and an antisense regulator domain. Antiswitches can be designed to activate or repress translation in response to ligand binding. ‘Off antiswitches’ feature an antisense domain that is stabilized as a stem loop in the absence of ligand, permitting translation of the targeted mRNA. Ligand binding to the aptamer domain triggers a conformational change, exposing the antisense domain and repressing translation. ‘On antiswitches’, on the other hand, feature an antisense domain that forms a stem loop when the ligand is bound, releasing the mRNA target from repression. Importantly, the aptamer and antisense domains are functionally modular; exchanging a theophylline aptamer domain for a tetracycline aptamer domain changes the specificity of an antiswitch to tetracycline without altering the antisense targeting.

The modularity of riboswitches has been exploited to construct logic gates (Win & Smolke 2008). Riboswitch logic gates incorporate ‘sensor’, ‘transmitter’ and ‘actuator’ domains (figure 2a). The sensor domain consists of a ligand-dependent aptamer, and the conformational change triggered by ligand binding alters the conformation of the transmitter domain. Attaching the sensor and transmitter to an actuator domain, in this case an HHRz, confers ligand-dependent ribozyme activity. The devices were embedded in the 3′ UTR of mRNA transcripts, such that HHRz activation triggers mRNA cleavage leading to decreased expression. More complex devices were constructed by linking pairs of sensor domains, linked by transmitters, to a single actuator. Rearrangement of dual sensor domains relative to an actuator domain yielded AND, NOR, NAND and OR gates (figure 2b).

Other devices were also constructed by placing two sensor–actuator pairs in series on an mRNA. One such device is a ‘bandpass filter’, so named because expression of the reporter protein only occurs at intermediate inducer concentrations, with low and high inducer concentrations leading to mRNA degradation. This was accomplished by including a ‘buffer gate’ that prevents ribozyme cleavage in the presence of theophylline as well as an ‘inverter gate’ on the same mRNA that activates ribozyme cleavage in the presence of theophylline.

Boolean logic gates have also been designed with small interfering RNAs (siRNAs; Rinaudo et al. 2007). In these devices, inputs trigger siRNA expression, with each siRNA targeting a specific mRNA. The targeted mRNA molecules each encode a repressor protein such as LacI that repress the expression of a fluorescent protein reporter. For example, an AND gate consists of two mRNAs driven by the same promoter, but with different 3′ UTR targets, A and B. Each mRNA encodes LacI or LacI–KRAB to repress the expression of the reporter protein. The presence of both siRNA A and siRNA B is required to knock down the mRNAs and allow reporter expression. An OR gate was constructed by expressing a single mRNA for the repressor protein, with two 3′ UTR targets, such that siRNA A or siRNA B is sufficient for knockdown.

This siRNA-based logic allowed the construction of devices that evaluate complex expressions such as ‘(A AND B AND C) OR (D AND E)’ and ‘(A AND C AND E) OR (NOT(A) AND B)’ (figure 2b), with A, B, C, D and E representing different inputs. Every possible input combination was experimentally verified, with the second device yielding the only incorrect output, returning false in the case that A and E were true and B and C were false (Rinaudo et al. 2007).

RNA logic devices demonstrate that RNA can regulate complex synthetic networks. Both the riboswitch and RNAi-based devices take advantage of antisense base pairing to confer specificity between synthetic RNA inputs and their mRNA targets. Modular riboswitches may prove to be portable to a wide range of host species, as RNA folding and ribozyme activity occur independently of the host cellular machinery. Complex RNA regulatory devices in natural systems, such as the S-adenosyl-methionine/adenosylcobalamin NOR gate in Bacillus clausii, demonstrate that RNA aptamers are an effective means of responding to changes in metabolite concentrations (Nahvi et al. 2002; Stoddard & Batey 2006).

In higher eukaryotes, introns appear to be an important regulatory element at the mRNA level. Approximately 95 per cent of the average human gene is made up of introns (Lander et al. 2001; Venter et al. 2001). Both introns and exons must be transcribed; introns are removed from the mRNA prior to translation. In the development of multicellular organisms, one suggested role for introns is in regulating the timing of gene expression, as intron processing increases the time between transcription and translation. Time delay coupled with autoinhibition can produce oscillations, and oscillatory gene expression is often observed in development (Swinburne & Silver 2008).

Evolution has found a solution to orthogonal signalling that still allows cells to use the same protein components for multiple processes (Bhattacharyya et al. 2006). Classes of proteins, such as kinases, share a common mechanism of action but can act on a variety of targets. This is often achieved by varying combinations of adapter domains and effector domains. Signalling protein interactions are often mediated by events such as phosphorylation that change binding affinities and ‘rewire’ the network (Pawson 2007).

Prokaryotic two-component signalling systems provide a simple model system for studying protein interactions. Canonical two-component systems consist of a membrane-bound receptor with ligand-dependent histidine kinase (HK) activity and a response regulator (RR) protein, usually a transcription factor (West & Stock 2001).

The rational design of two-component signalling systems has recently been demonstrated (Skerker et al. 2008). Multiple sequence alignments of HK and RR pairs identified residues that covaried, representing interacting partners. Modification of as few as three interacting residues in the HK EnvZ (an E. coli osmolarity sensor) switched the EnvZ HK phosphorylation specificity from its cognate RR OmpR to the non-cognate RR RstA, which is not normally induced by osmotic changes. Additionally, the RR specificity of EnvZ was switched to that of CC1181, a Caulobacter crecentus sensor protein. This research establishes a protocol for the rational design of two-component systems, as well as validating methods of HK–RR interaction prediction. The existence of two-component systems in eukaryotes (Saito 2001) implies that novel HK–RR pairs could augment eukaryotic devices as well.

Eukaryotic signalling cascades are usually more complex than two-component systems, but eukaryotic signalling proteins can still be reprogrammed to accept new inputs. For example, a guanine nucleotide exchange factor (GEF) was modified to be responsive to protein kinase A (PKA; Yeh et al. 2007). Active GEFs catalyse the exchange of GDP for GTP bound to Rho GTPases, and, in turn, GTP-bound Rho activates downstream effectors (Rossman et al. 2005). To build a synthetic GEF, researchers replaced the cognate autoinhibitory domain of the CDC42-specific GEF Itsn1 with a PKA-dependent autoinhibitory domain, leaving the GEF catalytic domain intact (figure 3; Yeh et al. 2007). The addition of forskolin, a PKA activator, induces production of filopodia in mammalian cells, indicating Itsn1 signalling. Substituting different GEF catalytic domains also produced new signalling behaviours, and signalling cascades involving two synthetic GEFs were also functional (Yeh et al. 2007).

Protein domains can be combined to produce novel switch-like behaviour. A chimeric protein with two separate ligand-binding domains could act as a switch or an OR gate if only one ligand-binding domain could be occupied at a time. To isolate new protein switches, researchers overlapped functional ligand-binding domains and peptides in chimeric proteins, such that correct folding of one domain would disrupt folding of the other (Sallee et al. 2007). Out of 25 candidates, seven chimeric proteins yielded functional switches, with domains that are unstructured in the absence of ligand showing the highest likelihood of success (Sallee et al. 2007).

Protein ligands that bind cell surface receptors can also be used as modular regulatory elements. One such device, a chimeric-activating protein, was constructed by connecting an epidermal growth factor (EGF) ligand and an interferonα-2a (IFNα-2a) ligand via a flexible linker (Cironi et al. 2008). The EGF ligand acts as a targeting element, binding the EGF receptor (EGFR). The IFNα-2a ligand triggers the desired action of the device, binding IFNα-2a–IFNα receptor 2 (IFNAR2) and activating the Jak–Stat pathway (Platanias 2005). In the chimeric activator, the IFNα-2a ligand was mutated to reduce its binding affinity for IFNAR2. Reducing IFNα-2a binding affinity had the desired effect: IFNα-2a-mediated activation of the Jak–Stat pathway occurs only when both EGFR and IFNAR2 were present on the cell surface. EGF binding to EGFR brings IFNα-2a closer to the cell surface, increasing the likelihood of IFNα-2a–IFNAR2 binding and subsequent Jak–Stat signalling. As well as having therapeutic applications, chimeric activators could be incorporated into synthetic devices for intercellular signalling (Cironi et al. 2008).

These synthetic protein devices demonstrate that the rearrangement of natural protein modules can yield new behaviours. To achieve protein–protein interactions that are truly orthogonal to an existing protein network, however, it may be necessary to design new protein interactions. There is evidence that new types of signalling, such as tyrosine kinase signalling, evolved in response to the saturation of previous signalling networks (King et al. 2003; Bhattacharyya et al. 2006). The engineering of novel signalling systems may permit synthetic devices to operate in cells without the interference of the endogenous protein network.

A significant effort in metabolic engineering has been the production of amorphadiene in E. coli (Martin et al. 2003) and artemisinic acid in S. cerevisiae (Ro et al. 2006), both precursors to the anti-malarial drug artemisinin. In both cases, the amorphadiene synthase (ADS) gene from Artemisia annua L was heterologously expressed, and flux through the host metabolic network was redirected towards ADS. In the case of S. cerevisiae, expressing the ADS is sufficient for amorphadiene production, albeit with low yields. Adjusting the expression levels of five genes involved in the production of farnesyl pyrophosphate (FPP, converted to amorphadiene by ADS) yielded a 500-fold increase in amorphadiene production (figure 4). Screening a library of A. annua cytochrome P450 expressed sequence tags yielded an enzyme that catalysed the conversion of amorphadiene to artemisinic acid, which was then integrated into the engineered strain along with NADPH : cytochrome P450 oxidoreductase. In the case of artemisinic acid, as well as in many other metabolic engineering efforts, the redirection of flux via adjustments in gene regulation was paramount in achieving commercially viable yields (Keasling 2008).

Re-routing metabolic flux via gene deletions can force cells to produce more of a desired product. Genome-scale simulations of cellular metabolism such as flux balance analysis (FBA), which we will explore in §6, can predict beneficial gene deletions. This approach has been validated for several metabolic engineering efforts, including the production of lycopene in E. coli (Alper et al. 2005), and in the case of our own efforts to produce formic acid in S. cerevisiae (Kennedy et al. in preparation). Results in both studies suggested that further regulatory modifications would boost yields. In silico strain design coupled with tight regulation of enzyme levels by synthetic devices will undoubtedly be essential to future metabolic engineering efforts.

As in protein signalling networks, modular manipulation of metabolic enzymes would facilitate the development of new pathways. However, unlike signalling proteins, metabolic enzymes are rarely structurally modular. For example, metabolic enzymes often contain allosteric regulatory sites and the active site within a single domain, confounding efforts to decouple allosteric regulation and catalysis (Bhattacharyya et al. 2006). A notable exception is enzymes such as polyketide synthases (PKS), which are not only modular, but their spatial arrangement can also be exploited to create new products.

Rational engineering of PKS assembly lines would open up new possibilities for the synthesis of organic molecules. Polyketides are assembled by linear complexes of PKS proteins, with each PKS performing catalysis on the growing polyketide chain (figure 5). Evolutionary rearrangement of PKS modules has generated a diverse array of natural products, including many antibiotics (Robinson 1991). PKS proteins possess N- and C-terminal ‘docking domains’ for attachment to other PKS (Thattai et al. 2007). Combinatorial shuffling of PKS modules resulted in the in vivo synthesis of novel polyketides (Menzella et al. 2005). Computational modelling of potential PKS products suggests that billions of possible molecules could be synthesized via engineered PKS combinations, and that it may be possible to predict PKS combinations that will produce a desired compound (González-Lergier et al. 2005).

Most constraint-based models of metabolism are based on the framework of FBA, a technique that simulates the entire metabolic network of an organism (Varma & Palsson 1994). The only required parameter for an FBA model is a stoichiometric matrix that contains all known metabolic reactions of an organism. Constraints are placed on certain fluxes, defining nutrient availability and relative uptake rates as well as thermodynamic constraints on the reversibility of reactions. It is assumed that, at steady state, the net flux of the system is fixed. The model is then solved for the optimization of an objective function such as maximization of biomass. Since FBA models do not consider enzymatic parameters beyond the stoichiometry of each reaction, the availability of comprehensive databases such as the Kyoto Encyclopedia of Genes and Genomes (http://www.kegg.com) has fostered the development of FBA models for many organisms (Varma & Palsson 1994; Duarte et al. 2004; Becker & Palsson 2005; Feist et al. 2007; Lee et al. 2008; Senger & Papoutsakis 2008). Owing to the genome-scale nature of constraint-based models, in silico screens have been applied to predictions of gene essentiality (Edwards & Palsson 2000; Thiele et al. 2005; Samal et al. 2006; Becker & Palsson 2008) and the related metabolic engineering problem of predicting gene knockouts for strain optimization (Burgard et al. 2003; Alper et al. 2005; Kennedy et al. in preparation).

The successful application of FBA in a variety of organisms demonstrates the use of constraint-based models in the context of metabolism. However, even in the unlikely case that enzyme kinetics are unimportant to determine metabolic flux, traditional FBA models assume that a cell's entire complement of enzymes is available at all times. Regulation can be modelled implicitly, via methods such as minimization of metabolic adjustment, which assumes that regulation will force mutant flux distributions to be as similar to the wild-type distribution as possible (Segrè et al. 2002). Models such as regulatory FBA attempt to explicitly model regulation by switching fluxes on and off, based on the experimental data of enzyme expression in various growth conditions (Covert et al. 2001; Covert & Palsson 2002; Herrgård et al. 2006).

Experimental efforts to modify transcription factor behaviour have underscored the importance of regulation to metabolic fluxes. One such approach, known as ‘global transcription machinery engineering’, yields improved strains by screening transcription factor mutants. For example, a more ethanol-tolerant strain of S. cerevisiae was isolated by mutagenizing the TATA-binding protein SPT15 (Alper et al. 2006). The global influence of transcription factors makes them a powerful tool for strain construction; unfortunately, this has also made it difficult to predict the full effects of transcription factor modification. Future genome-scale models may include genome-scale models of transcriptional regulation, but comprehensive information on transcription factor interactions is not presently available.

Metabolic regulation in vivo is influenced by enzyme and substrate concentrations and the kinetic parameters of each enzyme. Kinetic models sacrifice the genome scale of constraint-based models in favour of detailed quantitative modelling of specific pathways. Unlike FBA, metabolic control analysis (MCA) accounts for the kinetic parameters of all enzymes in the pathway, along with the concentrations of the enzymes and the metabolites involved (Fell 1997). Control coefficients for each enzyme, based on experimentally measured parameters, define the amount of influence an enzyme has on a pathway. A key concept of MCA is multisite modulation, i.e. there is no ‘master’ enzyme or rate-limiting step in a pathway; instead, each enzyme in the pathway has a non-zero control coefficient, and the control coefficients of all the enzymes in the pathway sum to 1. Thus, MCA can identify enzymes with high control coefficients, and quantitatively predict the impact of adjusting the concentration or rate of those enzymes. Unfortunately, the detailed nature of MCA models makes them impossible to apply on the genome scale until more comprehensive data on enzyme kinetics have been collected.

Even if ‘omics’ technologies generate a complete list of parameters for genome-scale MCA, the resulting dataset would still be difficult to model (Schuster 1999). One approach to reducing the complexity of a genome-scale model is to group metabolic fluxes into modules, provided that the groupings are functionally relevant. Although individual metabolic enzymes are rarely composed of modular elements, considering groups of enzymes as higher order modules may yield adequate models for understanding metabolic network kinetics at a genome scale.

Lysosomes, peroxisomes and many other organelles contain enzymes or metabolites that are harmful to the host cell (Page et al. 1998; Yeldandi et al. 2000). Metabolic reactions that would be thermodynamically unfavourable in the cytosol are often found in organelles (Feldman & Sigman 1983). Organelles could conceivably harbour engineered pathways that would be incompatible with the cytoplasm.

The specialized machinery of organelles could augment synthetic devices. Chloroplasts and mitochondria, being of endosymbiotic origin (Embley & Martin 2006), are the most sophisticated eukaryotic organelles and have useful properties for engineered systems. These organelles possess subcompartments of their own, the thylakoids in the case of chloroplast and the matrix in the case of the mitochondrion (Frey & Mannella 2000; Mustárdy et al. 2008). In addition, both organelles can generate electrochemical gradients between their subcompartments (Dimroth et al. 2000).

Although no completed examples of synthetic organelle devices exist, there is reason to believe that engineering organelles is feasible. In the case of many organelles, such as chloroplasts (Soll & Schleiff 2004), mitochondria (Truscott et al. 2003) and peroxisomes (Léon et al. 2006), the targeting of arbitrary proteins to these compartments has been described. Chloroplasts and mitochondria also have self-contained genomes, although gene integration into their genomes is more complex than nuclear integration (Bonnefoy & Fox 2007; Verma & Daniell 2007).

Co-cultures have been applied to the breakdown of lignocellulose, an important carbon source for biofuels (Eiteman et al. 2008). In current industrial processes, hydrolysis of lignocellulose yields a mix of five-carbon sugars, such as xylose, and six carbon sugars, such as glucose. In the presence of both xylose and glucose, E. coli will preferentially feed on glucose first. A co-culture of two E. coli strains, one that is deficient in xylose usage and the other that is deficient in glucose usage, breaks down the sugar mixture more effectively than a monoculture (Eiteman et al. 2008).

The synthetic devices presented here have demonstrated that functioning devices can be constructed, even though our understanding of biological systems is incomplete. In the most promising cases, engineering progress has also provided new biological insights. There is also much to be learned from building synthetic devices that do not work as planned. Designing synthetic devices in an iterative fashion, with experimental results allowing improved models and vice versa, will allow increasingly complex device designs. In addition, modelling shortfalls and unexpected experimental outcomes may shed light on new mechanisms of endogenous biological regulation.