The goal of creating a synthetic cell necessitates the development of reaction networks which will underlie all of its behaviours. Recent developments in in vitro systems, based upon both DNA and enzymes, have created networks capable of a range of behaviours e.g. information processing, adaptation and diffusive signalling. These networks are based upon reaction motifs that when combined together produce more complex behaviour. We highlight why it is inevitable that networks, based on enzymes or enzyme-like catalysts, will be required for the construction of a synthetic cell. We outline several of the challenges, including (a) timing, (b) regulation and (c) energy distribution, that must be overcome in order to transition from the simple networks we have today to much more complex networks capable of a variety of behaviours and which could find application one day within a synthetic cell.
Living cells have unique capabilities, such as self-healing, sensing and adapting to the environment, homeostasis, signalling, controlling gene expression and converting chemical energy into motion, ultimately allowing cells to grow and divide. On the basis of all these capabilities, a vast network of enzymatic reactions together forms the molecular wiring of the cell. These networks are capable of processing and storing information about the state of the cell and its surroundings in order to generate responses to optimise its behaviour(s). Exactly how this network functions, where its parts originate, and how such systems can self-organise in space and time, is not yet understood. Thus, one of the most fundamental questions in science is how a collection of inanimate molecules may be able to form a functional system that is capable of further evolution into a living system. In recent years, there has been a tremendous progress in our understanding of prebiotically plausible pathways towards the building blocks of life on earth and of Darwinian evolution of organisms . However, there is presently no convincing theory that would explain how simple molecules-formed structures or networks capable of functional evolution towards increasing functional complexity, ultimately arriving at a self-sustained, compartmentalised set of chemical reactions that we would recognise as being alive.
Today, the fundamental knowledge to answer questions related to the origin of life is lacking and only parts of these questions are likely to be answered in the decades to come. It is, therefore, appropriate to first focus on some of the most fundamental topics on how to create functional molecular systems. Systems that possess some of the features of their natural living counterparts, like the transfer of information or signal-transduction, while being dynamic and adaptive at different length- and timescales, and that function close-to- and far-from-equilibrium.
The vast cellular enzymatic reaction networks operate by controlling the concentration of molecules in space and time which is achieved by controlling the synthesis/activation and breakdown/inhibition of components within the network. In order to regulate the synthesis, mechanisms need to be in place that balance the rate of reactions within the system. Functional output in complex reaction networks requires large changes in the rates of reactions with minimal changes in ‘input’ (i.e. reactivity should be switched ‘on’ or ‘off’) within a narrow concentration range. Non-linear kinetics typically associated with autocatalysis, or ultrasensitive or allosteric responses, is thus essential for the creation of functional molecular systems. At present, only (sets of) enzymatic reactions demonstrate such a breadth of behaviour and thus, for now, it is inevitable that networks will require their usage. Ultimately, if we wish to find application for these networks in a synthetic cell which could act as e.g. bioreactors for chemical synthesis, then we must understand how to construct these systems from the bottom-up.
Inspired by the living systems, a variety of reaction networks have been reported that perform potentially useful behaviours, such as digital circuits for Boolean logic informational processing [2–4], oscillators, bistable switches and memories [5–9], neural networks  and dissipative self-assembling systems forming transient materials [11–16]. Many of the designs are based on so-called network motifs that are also found in many biological systems and typically involve a small number of positive and negative feedback loops. However, most of these approaches do not share common building blocks and potentially have incompatible reactions and conditions, an extreme example would be the redox chemistry and acidic conditions, required for systems based upon or akin to the Belousov–Zhabotinsky reaction, will be incompatible with networks based upon biological components like enzymes or DNA.
We focus here upon those networks, which could be realistically incorporated into a synthetic living system akin to a biological cell. Synthetic biology has made substantial strides in producing gene-based or synthetic metabolic networks and we refer the reader to several recent reviews [17–21]. However, we feel that an in vitro approach has much to offer, as the ability to control individual rates within the network, change environmental parameters such as temperature or concentration, or indeed systematically build the complexity of networks, allows a deeper molecular understanding into the behaviour and mechanisms that play a role in translating network design into a functional output. Examples of DNA, hybrid DNA + enzymatic and enzymatic networks are presented below.
In vitro DNA networks
Pure DNA-based synthetic reaction networks have proved to be some of the most fruitful networks [22,23] in terms of generating an impressive array of behaviour e.g. logic gates [24,25], digital circuits [3,26], oscillations , neural networks [10,28], chemical controllers for regulation , signalling networks . In these systems, an ‘input’ DNA strand initiates a toe-hold strand displacement with a complimentary strand upon which an ‘output’ DNA strand is bound (Figure 1a). The binding of the input strand releases the output strand which can trigger further toe-hold strand displacements in the network. The combination of several toe-hold strand displacements can be used to create Boolean logic operators e.g. AND gates (Figure 1b). Interactions with fluorescent probes are used to track the state of the network. Base pair encoding provides wide programmability and the high specificity of base pair interactions between complimentary strands provides the network with high precision. Furthermore, this allows for a diverse range of orthogonal reaction network pieces to be present in the same environment with few deleterious cross-reactions and provides low error rates. Such networks based upon toe-hold strand displacements typically operate on the timescales of several hours unless confined in a compartment, e.g. a proteinosome, whereupon this can be reduced to the minute timescale (Figure 1c) , or an acrobatic DNA is used which can reduce the timescale to seconds . DNA systems with deoxyribozymes  have also been developed and used to implement logic gates  and reversible logic gates .
Joint DNA and enzymatic networks have been developed including those based upon the Polymerase/Exonuclease/Nickase Dynamic Network Assembly (PEN DNA) toolbox  and also transcriptional genelet networks [7,35,36]. These networks use DNA to process and store information and use enzymes to synthesise and degrade the various DNA (and RNA) network components. The PEN DNA toolbox has been used to create bistable switches , signalling networks , oscillators  and predator–prey models . Meanwhile, transcriptional genelet networks have been used to develop oscillators [7,36] and clocks .
In vitro enzyme networks
As mentioned in the introduction, reaction networks function by controlling the concentrations of their components in space and time through the synthesis/activation and breakdown/inhibition of network components, which requires catalysts capable of responding in a non-linear manner in order to engender sharp transitions in concentration. At present, the only catalysts capable of such behaviour are protein-based enzymes. While nucleic acid-based catalysts like ribozymes exist, the scope of their reactions in natural systems is predominantly confined to amide bond formation and phosphate cleavage, meanwhile enzymes can catalyse at least 63 different sub-classes of reactions . Moreover, ribozymes’ greater flexibility, and, more limited and less varied scope of side chains, in comparison with enzymes, means that their potency to function as catalysts in the future is, relatively speaking, constrained .
Nature employs enzymes in reaction networks for a number of reasons: (i) Enzymes perform reactions over an enormous range of timescales up to and including the diffusion limit with kcat/KM values ranging from 100 to 109 M−1 s−1 . This allows for the construction of networks which function over a huge range of timescales from hours to the millisecond. (ii) Non-linearity may be introduced into enzymatic systems via multiple means including autocatalysis (e.g. in the activation of zymogens or in autophosphorylation reactions of kinases) for positive feedback loops, allosteric regulation for the control of conformation to affect catalysis, substrate competition for the active site or reactions/interactions with inhibitors and potentiators. (iii) Vast substrate scope, with new reactions, is constantly being evolved, either naturally [43,44] or by direction . Thus, networks based upon a wide array of different chemistries can be created.
Thus, Nature has interwoven nucleic acid-based networks with enzymatic reaction networks to create truly rich functional molecular systems. DNA is used to store and, to an extent, compute information, and the catalytic activity of RNA in the ribosome is obviously of great importance. However, Nature uses enzymatic reaction networks for processing information and short-term memory storage (predominantly via phosphorylation and dephosphorylation of proteins) in order to sense (and communicate with) the environment, to gain self-healing properties, to introduce motion, to extract energy out of nutrients and to synthesise life's building blocks.
In our work, we have set out to design, build and test the functional output of out-of-equilibrium chemical reaction networks of increasing complexity. We have used enzymes for the construction of complex joint enzymatic and small molecule networks [8,46–48]. We initially wished to design a network capable of oscillatory behaviour and as a first step, we chose a suitable oscillating network topology based upon an autocatalytic positive feedback loop and a delayed negative feedback loop (Figure 2a). A retrosynthetic approach was then employed where the different nodes in the network (enzymes) were connected via small molecules acting as substrates (Figure 2b). In this network, the positive feedback loop trypsin (Tr) cleaves its own zymogen trypsinogen (Tg) in an autocatalytic reaction. In the delayed negative feedback loop, first activation takes place where a protected inhibitor (Ac-Lys(Me)-Gln-Inh, 2) has the first amino acid residue cleaved by trypsin to reveal an intermediate inhibitor (H-Gln-Inh). In the delay step, the intermediate inhibitor is cleaved further by an orthogonal enzyme, aminopeptidase (Ap), to reveal the trypsin inhibitor (Inh) which then irreversibly inhibits trypsin in the inhibition step. Oscillations in trypsin concentration are produced due to switching between dominance of the positive feedback loop and dominance of the negative feedback loop over time. If the strength of the positive and negative feedback loops become equivalent, then a steady-state concentration of trypsin is produced. All rate constants were determined from isolated individual reactions using purified components (Figure 2c). Crucially, the system was then assembled in a flow reactor to maintain out-of-equilibrium conditions (Figure 2d). Finally, the output of the reactor was measured using an enzymatic activity assay or determined using chromatographic methods (HPLC). This network was simulated solving sets of differential equations for each individual reaction and comparison to experimental results yielded a good fit with the replication of the periodicity to within ±0.2 h (Figure 2e).
Design and implementation of an oscillating enzymatic reaction network.
The strength of our approach lies in the ability to exploit the full power of chemical synthesis to tune the behaviour of the networks by small molecules. To exemplify, Figure 3a shows a number of derivatives of 1 (one can easily envisage a much larger number of functional groups included) and partially unpublished work where we have measured the rate constants of activation (kact) and inhibition (kinh) in the negative feedback loop. Different substituents in R1 enable fine-tuning of activation since the N-terminus of the pro-inhibitor will influence the binding affinity to trypsin. We have recently reported how R1 leads to subtle changes to the amplitude, periodicity and parameter space (flow rate, enzyme and pro-inhibitor concentrations) where limit-cycle oscillations can be found . Different substituents in R2 will give an increase in inhibitor rate constant when going from methyl to propyl (Figure 3b), and show an interesting and unexpected increase in robustness of the response, where a smaller change in inhibitor concentration leads to a sharper drop in trypsin activity over time (Figure 3c). We refer readers to the original papers for a detailed discussion of these effects.
The effect of variations in the chemical structure of the protected inhibitor on the trypsin oscillator network.
Beyond oscillations, we have also developed reaction networks based upon feed-forward loops for adaptive response behaviour (Figure 4a–c)  and bistability (Figure 4d,e) . The adaptive response network detects a change in the environment and triggers a response to this stimulus. After registering stimulus, the network then resets itself back into its original state whereupon a second stimulus can trigger the network response again. We built this network with a feed-forward loop based upon the enzymes trypsin and chymotrypsin (Cr) (Figure 4a). A small peptide Z-Phe-Arg-AMC can be cleaved by both trypsin and chymotrypsin but at different positions. If trypsin cleaves the peptide, then it releases a fluorescent AMC. However, if chymotrypsin cleaves the peptide first, then the resulting Arg-AMC is no longer a substrate for trypsin and no fluorescent AMC is released. In the absence of trypsin, no increase in fluorescence is observed but upon the addition of a trypsin stimulus, a spike in fluorescence is seen (Figure 4b,c). The fluorescent signal is reset by trypsin-activating chymotrypsinogen (Cg), the zymogen of chymotrypsin. The increased chymotrypsin concentration then outcompetes trypsin for cleavage of Z-Phe-Arg-AMC causing a loss in fluorescent signal and resetting the network. In our bistable network (Figure 4d), we create a positive feedback loop with trypsin and trypsinogen and combine this with a strong inhibitor of trypsin (4-(2-guanidinoethyl)benzene-sulfonyl fluoride, GEBSF). When placed in a flow reactor, the network can show bistability for a set of space velocities (light green region). Whether the network is either in low or high trypsin concentration depends upon the history of the system (hysteresis) (Figure 4e). If the network is in a high trypsin state, then autocatalysis of trypsin can outcompete its inhibition by GEBSF and vice versa for low trypsin state. Outside the bistable region, the inhibition is either too weak that there is a high trypsin concentration (yellow) or too strong and thus there is a low trypsin concentration (dark green).
The central challenge going forward for reaction networks in synthetic living systems is how do we transition from our current set of relatively simple complex systems to far more complicated complex systems capable of a range of behaviours that could one day find the real-world application. At present, we are able to construct a range of different simple modules that can act as e.g. oscillators and switches, but as we develop more complex reaction networks, we will need to learn how to arrange the different components and motifs to optimally function together. Here we outline several medium-term challenges that need to be addressed in order to produce more sophisticated networks: (a) timing, (b) regulation and (c) energy distribution.
As the key to any good joke, timing is of utmost importance in reaction network design. As networks become more elaborate, timing the output signal(s) generated upstream so that all signals into a downstream receiver coincide, will be required in order to produce the desired behaviour. In particular, for processing information, this must be done in the correct order which requires control over timing. Such control can be exercised by adjusting the rates of reaction within the network of which there are four possibilities for doing this: (i) controlling the concentrations of reagents; (ii) affecting the magnitude of the rate constant through modification of the chemical structure of the substrate; (iii) environmental control from adjusting e.g. temperature or via exposure to light;  (iv) compartmentalisation and crowding [51–55]. The optimal balance of rates for particular behaviour can thus be achieved via adjusting either one or a combination of these factors.
At the upper end, reaction rates are constrained by the diffusion limit and if such a reaction is utilised then every other part of a reaction network will need to be arranged relative to this hard limit. Arguably, the most critical reactions and motifs, such as those which provide energy to keep the system out-of-equilibrium or enzymes which are responsible for maintaining a stable environment, should operate either at the diffusion limit, or as close to this limit as possible, as not every enzyme reaction will be capable of reaching the diffusion limit due to the chemistry involved. This is indeed the strategy that Nature has adopted as ‘perfect' enzymes, which operate at the diffusion limit, are employed in these roles, e.g. DHAP/GAP TIM etc which ensure that glycolysis produces a net production of 2 ATP and carbonic anhydrase which maintains a constant pH . However, operating all reactions in networks at the diffusion limit would likely lead to a breakdown in the network due to an unsustainable drain on resources. Thus all other processes (e.g. signalling, information processing etc.) will need to operate on timescales below the diffusion limit, with coordination of motif behaviour in time required in order to keep networks functional.
The regulation of network behaviour is vital in order to be able to produce consistent behaviour when subject to a range of different environmental conditions. Thus far, we have shown that our approach can be successfully applied to construct small enzymatic reaction networks with well-defined functionalities within a given environment. However, in Nature, the environment is hardly ever stable, in contrast, with the model systems, we study in the laboratory. Although the limit-cycle oscillations of our trypsin oscillator are robust over a fairly large phase space (i.e. combination of reaction conditions such as temperature and flow rate as well as starting concentrations of enzymes or small molecules), any change in these conditions will change the periodicity of the oscillations! This dependency of the precise functional output prohibits future applications. For example, oscillators would obviously make an interesting clock for a synthetic cell, but if the clock changes pace with every fluctuation in salt concentration, temperature or volume of the cell, then it will be impossible to use this to provide temporal control over the system as a whole. In Nature, regulation can be achieved through either resisting environmental changes via the incorporation of feedback mechanisms, or, by embracing the environmental changes through the synchronisation of behaviour with the environment e.g. circadian rhythm [57,58,59].
Energy and resource distribution in reaction networks has not so far been an issue for in vitro networks due to the reliance upon simple designs where there is little to no competition for resources. Competition for resources has been exploited to produce networks that utilise a winner-takes-all decision making process [28,60,61,62]. However, the challenge here is how to keep a complicated complex system, which contains multiple parallel sub-networks that share resources, operating in unison. A winner takes all approach here is not desired as this would result in a complicated network simplifying itself into only a few sub-networks. The problem of distributing energy and resources is a well-known challenge in metabolic engineering where mismatches in energy demand in synthetic metabolic pathways severely affect cell function, even leading to cell death [63,64]. In a synthetic cell, the challenge will be to ensure that parts of the network that use the same energy sources and resources, operate upon the same timescale and that those parts of the network which are responsible for energy generation operate at a faster timescale than those processes which consume energy in order to be able to match demand. Inventing new means of regulation, inspired by processes that regulate enzymatic activity as mentioned above or subcompartmentalisation of certain processes, will be crucial. Ultimately, a network that can produce and regulate energy distribution will form the central spine to any synthetic living system.
To solve the above three challenges, we require tools which can predict the behaviour of combined components. Taking a clock as an example, we know how to build an oscillator but even in the simplest case of oscillatory motifs, we can only design the system to oscillate, but there is no known route to a priori design the periodicity (which is set by the relative rate of the reactions and the flow rate). This is a challenging problem that cannot be solved purely by the rational design. Here, we take heart from recent advances in computer science and artificial intelligence. Evolutionary algorithms can be employed that will broaden an initial motif with additional feedback loops of different reaction strength, with the goal of producing not only oscillations, but also oscillations of maximum frequency . These algorithms test the capacity of networks to oscillate and select those with the maximum frequency. They then create ‘offspring’ by adding additional feedback loops, ‘mutate’ reaction rates, and again select for oscillators with maximum frequency (or go back if no more optimal networks are generated). In this way, new networks will be generated that could not be designed, as there are no design rules. Thus far, these approaches have only been shown to function in silico but given our ability to connect a range of different enzymatic reaction networks via small molecules, and tune the rate of reactions via the molecules structure of these molecules, we are confident that future networks based on evolutionary designs could be built and tested. A key challenge for this approach will be how to ask the right question of the algorithm in order to obtain a realistic and implementable network. Algorithms will generate an answer, but there are considerations that algorithms which are based only upon kinetics will be unable to account for, e.g. solubility of networks components, and thus modelling will require careful boundary settings to avoid these pitfalls. Another challenge for this approach is that it will require large datasets of enzymatic reactions. Fortunately, databases with rate constants for enzymatic reactions already exist [66,67], although means by which this information may be parsed to identify substrates which are compatible with every reaction in the motif/network will need to be developed.
Beyond identifying networks which can produce behaviour with specific properties, we also wish to develop a richer range of behaviour from reaction networks. This will require the coupling together of many different reaction motifs . In order to develop such increasingly complicated systems, we require the chemical reactions that comprise these motifs to be compatible with one another i.e. are sufficiently orthogonal in reactivity to avoid deleterious side reactions [69,70]. Unfortunately, much of chemistry that has been developed over the last 150 years is incompatible for the design of networks beyond a single network motif. Chemists have adopted a linear mindset when developing chemical reactions which proceeds from starting reagents, through intermediates to products with purification at each step if necessary. Most small-molecule catalysts developed for this approach are essentially just active sites and control is, at most, exercised up to the second coordination sphere. Such a manner of thinking produces catalysts that are incompatible with complex network formation as they will be highly vulnerable to side reactions which either lead to undesired products or which deactivate the catalytic center. This necessitates the use of enzymes for the construction of complicated networks as the bulk of an enzyme structure protects the active site and prevents side reactions. Integration of non-biological chemistry may be possible through artificial enzymes [71,72], scaffolding  and compartmentalisation. Although there is a limit to compartmentalisation as if every sub-network is placed in an individual compartment, then as the complexity of the network has increased, the number of compartments required will quickly become absurd and the challenges such as communication between compartments will arise.
Finally, as networks become increasingly complex, we will need to begin to ask the question of what is the overall purpose and priorities of the network are. The arrangement of timing, regulation and energy distribution will need to be aligned in order to maximise the network's purpose. This may sound a bit nebulous, and is perhaps illustrated by an example: in cells the arrangement of the metabolic network is suggested to be guided by the principle of maximising the growth rate with respect to an upper limit upon the Gibbs free energy of dissipation . Consideration of thermodynamic factors for out-of-equilibrium systems such as this or dissipative adaptation [75,76] may be necessary to understand how best to construct networks. Although how precisely these principles connect with the individual reactions that comprise the network is still far from clear.
The networks discussed here and constructed thus far are not even remotely close to self-sustaining networks that could power synthetic cells to grow and divide. However, they form the first steps towards a type of molecular information processing system or molecular computer that could start to act as a regulatory network to control certain processes in a synthetic cell. Initially, we will build functions such as clocks and switches, and learn how we can time these functions. A major next step will be to develop a network that regulates energy supply. Once this has been achieved, we will need to learn how to ‘plug in’ different sub-networks into this energy supply, without disrupting all other sub-networks. And finally, of course, this will all have to be achieved within the synthetic cellular compartment which is able to harvest energy and resources from its environment and expel waste. It is clear that the construction of a synthetic living system is one of the greatest outstanding scientific challenges.
DNA and Enzyme-based in vitro reaction networks have been demonstrated to produce a wide array of functional behaviours. However, the networks we are currently able to design are simple compared with those in extant life. The central challenge for future reaction network design is how to construct far more complicated complex systems.
In the development of more sophisticated networks, close attention must be paid as to how to optimise the timings of interactions between various parts of the network, how to regulate the network to ensure consistent behaviour and how to distribute energy and resources amongst all the network components.
Nature has evolved enzymes with various properties (e.g. catalysis over a wide range of timescales and ability to perform non-linear reaction kinetics) that make them ideally suited for incorporation into more complex networks.
Few catalysts developed by chemists are well suited for incorporation into complex reaction networks. Thus, for now, it is inevitable that enzymes or enzyme-like catalysts will be required in order to develop reaction networks for a synthetic cell.
To create more complex systems many different sub-networks need to coexist with one another. Compatibility between all components is thus required, or, if not compartmentalisation must be used.
The Authors declare that there are no competing interests associated with the manuscript.