Why cellular computations challenge our design principles

Abstract

Biological systems inherently perform computations, inspiring synthetic biologists to engineer biological systems capable of executing predefined computational functions for diverse applications. Typically, this involves applying principles from the design of conventional silicon-based computers to create novel biological systems, such as genetic Boolean gates and circuits. However, the natural evolution of biological computation has not adhered to these principles, and this distinction warrants careful consideration. Here, we explore several concepts connecting computational theory, living cells, and computers, which may offer insights into the development of increasingly sophisticated biological computations. While conventional computers approach theoretical limits, solving nearly all problems that are computationally solvable, biological computers have the opportunity to outperform them in specific niches and problem domains. Crucially, biocomputation does not necessarily need to scale to rival or replicate the capabilities of electronic computation. Rather, efforts to re-engineer biology must recognise that life has evolved and optimised itself to solve specific problems using its own principles. Consequently, intelligently designed cellular computations will diverge from traditional computing in both implementation and application.

1. Introduction

The modern electronic computers that we intelligently design and build have computational power that in practice borders upon the supposed theoretical limits of computation. By this we mean that more or less all of the computational problems that are solvable can be solved by our computers, at least in theory. Living cells solve problems too, and very successfully, although of a different class, or at least in a different domain. This despite living cells looking almost nothing like our own computing machines [1]. Cells must switch metabolic pathways depending on the presence of substrates. They must control their life cycle, optimizing strategies to divide or enter dormancy. They organize communities depending on the environmental conditions. They can move as a function of Earth’s magnetic field [2]. They build multicellular organisms. How do they solve these problems? This question is not about superficial differences between the physical implementations of computers versus cells, which we can attribute to differences in material substrate and environment. Instead, we wonder why the fundamental principles that we use to design computers are not obviously or reliably present in living cells, especially if they are so successful at encoding information and computing.

The observation that the operation of biological processes resembles computation encourages synthetic biologists to construct biological systems that perform predefined computations [3], with applications ranging from biomedical [4] to environmental [5] to the generation of smart-materials [6]. In many of these cases synthetic biologists are applying principles used in the design of conventional computers to the design of novel biological ones. But if it turns out that the evolution of biological computers has not used these same principles, it should matter deeply why, because some explanations might hint at constraints that also apply to the biological computers that we intelligently design.

For example, computer scientists and engineers know that a small set of simple operations called primitives can be connected together and reused to perform arbitrary computations within a given paradigm. Famously, all possible digital logic circuits can be built only by using connected NAND logic gates,² and using this principle, we have largely automated the design of wonderfully powerful computers from just a few distinct, cheaply manufactured and highly optimised components.

This same approach has been used in synthetic biology to design and build genetic logic circuits in living cells [7]. These are synthetic gene networks built from a set of primitive genetic logic gates (in this case NOR gates and NOT gates), that can be connected together in increasing number in order to solve increasingly large computational problems. There are examples of tremendous successes of this approach in solving small computational problems, most commonly, in biosensing domains [8]. Needless to say, however, genetic logic circuits have achieved nowhere near the scale of their electronic counterparts, with orthogonality, modularity, burden and robustness the most cited issues [8,9].

But why are these issues in the first place? For instance, if orthogonality prevents building genetic logic circuits for solving big computational problems, why didn’t living cells evolve some kind of more specific wiring system for genes? If one takes the position that there are extremely insoluble and fundamental problems that prevent Nature from addressing these kind of issues, it is hard to be too optimistic about the timeline for solving these problems ourselves. On the other end of the scale, one might argue that Nature hasn’t addressed these issues because building genetic logic circuits is simply not the best approach, that complex biocomputations are evolved using a completely different set of principles, and that we might one day discover and exploit those principles to design our own biocomputers far more powerful than gigantic genetic logic circuits.

The truth probably lies somewhere between these two positions. We would like to point out that thinking about not just the what, but also the why of the differences between our computers and living cells could be a lot more productive for figuring out how to apply what we know about the former, to the latter. Although we don’t yet have definitive answers, here we review several ideas that relate the theory of computation, living cells and computers, and suggest where these relationships might hint at possible solutions and pathways forward for the intelligent design of increasingly complex biological computation.

2. Why are living cells complex?

A Turing machine is an abstract mathematical model of a computer. The study of these machines is a foundation of Computer Science, and produces some deep insights into the act of computation itself, one of which is expressed as the Church-Turing thesis. A very informal statement of the thesis is: Everything it is possible to compute can be computed with a Turing machine.

This tells us that Turing machines are very powerful, but not uniquely powerful. There are many models of computers that are Turing-complete, which means they are equivalent to Turing machines, and therefore just as capable. Although at first glance many of these do not look like a Turing machine at all, we know of many very different models with exactly the same capacity for computation as the Turing machine. Interestingly, the threshold of complexity for Turing-completeness is very low, and some surprisingly simple models can be proved equivalent to Turing machines [10], and can therefore compute everything that it is possible to compute. Contrasted with these model computers, living organisms are extremely complex systems. Despite this, we recognise that they are still computers [11]; they compute by taking inputs from their environment to make decisions about how to invest their resources toward some objective, which can broadly and generally be described as maximizing their “fitness”.

This ability to adapt to a dynamically changing environment might be one of Life’s fundamental features, and these computations are performed by a biological machinery that has been evolving for billions of years since the Origin of Life [12]. However, no clear or unique set of principles appears to govern biological computations. Exploring the repertoire of decision-making mechanisms in biology reveals that their implementation spans multiple levels and operates within diverse paradigms. For example, gene regulation does not rely solely on the action of transcription factors, but is also controlled by a myriad of other phenomena, such as the effect of epigenetic signals [13], the spatial location of chromatin [14] and other cellular machinery [15,16], or the cellular context [17]—among others (Fig. 1C).

Fig. 1. Living implementations challenge traditional frameworks.

A. A conventional AND logic gate processes information linearly, from inputs to outputs (top). However, less intuitive designs can result in intricate regulatory networks (bottom) that maximise computational performance in living cells. B. An OR logic gate can be implemented in a straightforward linear configuration (top), but natural regulatory networks performing OR logic frequently exhibit highly complex physical structures (bottom). C. Unlike conventional silicon-based computing devices, living implementations are inherently complex systems, where circuits interact with other cellular machinery, which is further influenced by population dynamics. (Illustration of the cell inside by David S. Goodsell, RCSB Protein Data Bank. doi: 10.2210/rcsb_pdb/goodsell-gallery-028; population picture from a previous work on colony development [20]).

Moreover, information flows in biological systems often deviate from conventional standards. For instance, the implementation of a genetic AND logic gate might intuitively be expected to follow a straightforward, linear flow from two inputs to one output (Fig. 1A, top). However, computational optimisation of such a network suggests a highly intricate circuit with a counter-intuitive architecture [18] (Fig. 1A, bottom). Similar patterns can be observed in natural systems. For example, the typical way that synthetic biologists implement an OR logic gate is depicted in Fig. 1A, top. By contrast, naturally occurring OR logic gates can be much more complex. For example, in the TOL network of the soil bacterium Pseudomonas putida, shown in Fig. 1B (bottom), can perform an OR function [19]. The activation of the metabolic pathway encoded in the network is regulated by the XylS receptor, which is activated by either the presence of the pathway substrate (m-xylene) through a regulator protein (XylR), or directly by pathway intermediate products (3-methyl-benzoate). However, the TOL network is a significantly more complex structure than the OR-gate depicted in Fig. 1B, top — probably due to hidden interactions within the cell.

These diverse structures and mechanisms within biological systems produce emergent behaviour that is extremely difficult to understand and predict, even if the behaviour of components of the system might be well understood individually. This is what we mean by complex in this context. But their complexity does not help living cells solve more computational problems than a Turing machine. Firstly, the Church-Turing thesis suggests that living cells are not (and never will be) actually capable of solving any problem that a Turing machine cannot. Second, the existence of incredibly simple Turing-complete models of computation [10] suggests that complexity is just not necessary for biology to compute more things.³ Furthermore, it can be shown in a wide variety of settings that the kinds of chemical reaction networks upon which living systems are based are at least as powerful as Turing machines [21–23]. The conclusion we can draw is that biological computers are equivalent to Turing machines, and to conventional computers, in terms of their theoretical ability to solve computational problems.

It should be remembered that there is no mathematical proof, nor refutation, of the Church-Turing thesis, so its status and the details of its interpretation are still being litigated amongst mathematicians, logicians and computer scientists [24]. Because of this, and because the thesis is presented with reference to abstract models of computation like Turing Machines or Lambda calculus, it might be tempting to regard the notion of Turing-completeness as purely of philosophical or mathematical interest. But part of Turing’s genius was to establish the direct connection between the abstract world of mathematical models of computation and the physical world of mechanical computing machines, and in that context the thesis has real world consequences for both intelligently designed and naturally evolved computers. One of these is that the complexity we see in living cells is not a prerequisite for computational capacity, and that a more complex cell is not inherently capable of computing things that a simpler cell, or a conventional computer, could not. Living cells have not evolved complexity to be able to solve more computational problems.

3. Computational complexity

Computer science has developed its own notion of complexity, computational complexity, which has to do with how fast the time or space needed to solve a problem grows with the problem size. A Turing machine which solves a given problem has a time complexity which is the number of steps it takes to arrive at a solution, as a function of the size of the problem. Similarly, it has a space complexity which is the amount of storage space it uses to arrive at a solution, as a function of the problem size. A computational problem has space and time complexity, which are the corresponding complexities of the best (least steps, least space) Turing machine for solving the problem. If the complexity of biological computing systems has nothing to do with the power of those systems to solve more computational problems, could biological complexity be more closely related to computational complexity?

There are already some results related to this kind of biocomputational complexity. The usual approach is to choose a mathematical model of the biocomputational network, show how this mathematical model is equivalent to a computational model like a Turing machine or a logic circuit, then use the standard measures of computational complexity to characterise the biological system [25–27]. More results can be arrived at indirectly by a cross-disciplinary look at the literature, by first establishing different mathematical models for various biological networks, for example, Markov decision processes [28] or P-systems [29], then using the theoretical results on the computational complexity of these models [29,30] to infer their biocomputational complexity.

There may be objections that these approaches to characterising computational complexity are measuring things that are largely irrelevant to biological systems. However, space complexity has an intuitive mapping onto biological systems, and real consequences. When Leonard Adleman used DNA strands to solve an instance of the Hamiltonian Path problem [31] and give birth to the field of DNA computing, he used DNA strands as space. In doing so he also provided a great example of the consequences of space complexity for a biocomputation. Finding a Hamiltonian path is a computationally hard (in computer science jargon, it is an NP-complete) problem, and the solution used by Adleman worked by first generating all possible paths, encoded in DNA sequences, and then filtering out those paths that were not solutions.

The amount of space (DNA), that is required for this computation depends on the number of possible paths, and the number of paths grows very quickly as the problem gets bigger, so we can say that the space complexity of this biocomputation is very high. The impact is catastrophic and inevitable, and even for modest problem sizes (a couple of hundred vertices) the sheer amount of DNA required (weighing in at around six earths) pushes the computation completely beyond what is possible in practice. The point is not that this remarkable experimental achievement is of no use or interest, but rather that despite its ingenuity it still cannot escape the unhappy consequences of space complexity theory. And this remains true of subsequent examples of DNA computing that solve similarly hard problems [32].

Of course, the measurement of time complexity in units of “steps of a Turing machine” somehow seems much less relevant when talking about biocomputers. We sympathise with this viewpoint, but would like to argue that would be imprudent to disregard time complexity just because it is not intuitively related to time in biological systems. Famously, time and space complexity have a relationship, in that computational solutions can often trade space for time and vice versa. There is no reason to believe that this relationship doesn’t also hold in biological systems, and so the trade-off between time and space might be a useful design tool for both naturally evolved and intelligently designed biocomputations.

Indeed, time, and time complexity, may be much more important to biocomputations, because their biological inputs and outputs have their own characteristic timescales. Expression levels of a particular protein take time to respond to the concentration of some input molecule — the biocomputation downstream will have to wait. But not too long, because proteins are also degraded and diluted — the biocomputation must finish before then. The steps of a biocomputation are therefore subject to upper and lower time bounds that depend on the timescales of the biological processes implementing them. To complicate matters further, these timescales can vary by several orders of magnitude [33].

Overall, computational complexity places the same kinds of constraints on biological computers as conventional ones. The implementation of biocomputers must deal with a space and time complexity that is intrinsic to the computational problem being solved. Being aware of the complexity of different computational solutions, the physical limits imposed by biology, as well as the trade-offs that exist between them, will help push biocomputations into the domain of problems with greater computational complexity.

4. Information theory constrains both computation and biology

Information theory is a mathematical framework which deals with what information is, how it can be quantified, and how it can be communicated through channels. These are fundamental factors that limit and enable computation of all kinds. For example, in the Turing machine depicted as a graph in Fig. 2A, the nodes (circles) represent states of the Turing machine. Each of these states contains information about the previous states, and this information flows along channels (arrows) to subsequent states. The way we can conceptualise biological networks is strikingly similar. In Fig. 2B, a representation of the central carbon metabolism of Pseudomonas putida is also depicted as a graph, this time with the nodes representing metabolites, connected by channels that are actually biochemical reactions.

Fig. 2. From information flows to biocomputable problems.

A. A Turing Machine represented as a state diagram, with an initial state q₁ and an accept state q_accept. Information flows are shown as arrows connecting nodes (adapted from Figure 3.5 of Michael Sipser’s Introduction to the Theory of Computation). B. A schematic representation of the central carbon metabolism in the soil bacterium Pseudomonas putida, illustrating the pathway from glucose and gluconate to pyruvate. Information flows, analogous to those in the Turing Machine, are shown as directed arrows between nodes. However, unlike the Turing Machine, this system features numerous interactions (depicted here as red shapes) with the broader cellular machinery, leading to complex unpredictable dynamics. C. Every computable problem has a Turing machine that can solve it. In principle, there is also a biocomputer, a living cell for example, that can solve the same problem.

The key distinction lies in biological complexity. In the case of the Turing machine, all the states and channels are explicitly and completely captured. By contrast, it is impossible to create a representation that fully encapsulates all of the states and channels in the metabolism of Pseudomonas putida: hidden interactions within the network, with the broader cellular machinery and with the environment contribute to a complex behaviour that cannot be mapped out comprehensively. For instance, using the diagram in Fig. 2B, it would be hard to appreciate the complex relationship between sugar catabolic pathways in Pseudomonas putida and the redox balance, where different isozymes could act on different physiological conditions [34,35]. In addition, these interactions typically have relatively high levels of both extrinsic and intrinsic noise.

Nevertheless, we can still use this conceptualisation to apply results from information theory to biology, and in particular the link between information theory and molecular biology has a long history [36]. The central dogma of molecular biology describes the flow of genetic information through the processes of DNA replication, DNA transcription and RNA translation. Analysis of these processes reveals ways in which biology has optimised information theoretic properties, but also ways in which biology is suboptimal from a information theoretic point of view. For example, the existence of exactly four base pairs in DNA can be explained as an evolutionary optimisation of the information capacity of the DNA replication “channel” [37]. By contrast, a similar analysis of the genetic code mapping codons to amino acids reveals a suboptimal information capacity, even though a small number of changes could increase the capacity of this genetic code “channel” [38].

This naturally gives rise to questions. If the purpose of DNA is to communicate genetic information across generations, why would it have evolved to do that with a genetic code which is suboptimal according to information theory? Perhaps there are physical constraints which would explain or prevent further optimisation by evolution (and by us) in this direction. Of course, the complexity of biological systems means that genetic information is not the only type of information in a living cell, and also that there are many more information processing “channels” than the central dogma allows for. In biological systems, sub-systems can influence each other, leading to information being distributed across a much broader range of components [39]. At the same time, the nature of the molecular media imposes fundamental constraints on the amount of information that can be stored and transmitted [40,41]. As such, the problem of optimising information capacity across entire living cells is no doubt a much more difficult problem than across any single “channel”.

One open question is just how much information a cell can process. Studies have attempted to use the mutual information between inputs and outputs to estimate the information processing capacity of larger biological networks. For example, Cheong et al. [42] concluded that the immune cells they studied were probably only just about capable of Boolean (yes or no) decisions. However, later research has indicated that the mechanisms through which cells process information might use strategies to increase capacity, such as integrating the signals over time [43]. Moreover, previous reinterpretations under other theoretical frameworks [40] or experimental techniques [44] have shown a more optimistic scenario where the same immune cells can process more information. To our knowledge, the same measurement has not been reported in human designed biocomputation circuits.

Information travels from one system to another through channels, and the channel can limit the maximum amount of information able to be transmitted. In silicon-based computing devices, the channels are wires engineered to preserve most information between systems (e.g., memory, CPU, routers, etc). In biology, molecular interactions enable specific signalling inside and outside the cell, connecting systems in a similar way to wires in electronics. However, transmitting information using these interactions presents at least three challenges. First, the molecular interactions are sensitive to relatively high levels of noise, affecting the rate and stability of molecular interactions. Second, although most of the fundamental aspects of these molecular interactions are well known, understanding and predicting them for large molecules in solution is challenging. Finally, unexpected or unintended interactions often occur, which produce unwanted effects such as crosstalk or toxicity [45,46].

The first issue can be addressed by carefully considering the architecture and the medium in which biocomputation takes place. A helpful distinction consists of splitting the medium into intracellular and extracellular levels. At the intracellular level, the effect of molecular condensates [47] and co-occurring phenomena (e.g., translation [48]) can have a crucial role in reducing cellular noise; and there are strategies to digitalise signals [49], helping to further decrease their sensitivity to noise and maintain information stability. Therefore, intracellular spatial aspects are increasingly relevant for modelling [50,51]. At the extra-cellular level, the degradation of information over long distances becomes a concern, so different mediums of signal transmission might be considered depending on their asymptotic behaviour [41]. The careful location of bio-computational components that communicate through chemical signals could become a tool for circuit design [52]. If more considerable distances need to be considered between components, using electrogenic signals exhibits a smoother decay with distance than the diffusion of chemical signals [41]. Previous work has indicated its potential for bio-computation [53] or bio-electronic hybrid computation [54]. A final possibility is the design of complex spatial devices through advanced micro-fabrication techniques, enabling graph-like structures that arrange communication through specific three-dimensional channels [55].

The second and third issues can be addressed by carefully selecting components that both perform the desired function and minimize burden or toxicity. Fortunately, nature offers a vast repertoire of these components which we can mine [56], optimize [57], and select to fit our requirements best. Still, this approach constrains biocomputation to those components evolved by nature, regardless of our own requirements. Alternatively, these issues could be addressed through molecular engineering. A straightforward approach generates interactions between molecules (DNA, proteins, ligands) with high affinity and specificity. The engineering of hydrogen bonds in helix-bundles has enabled the building of very specific protein-protein interactions from simple design principles [58] which led to a large collection of protein-interaction based logic circuits. The modular combination of protein domains also enables powerful designs. In [59], DNA-binding Zinc-finger domains were combined with dimerizing domains (GCN4) and ligand-induced dimerizing domains (FkBP) to create multistable systems. Additional theoretical work based on the realization of the abundance of many-to-many protein-protein interactions [39] in nature has proposed the usage of protein-protein interactions-based biocomputing [60]. In this line, the recent progress on computational protein design [61–64] is accelerating the development of specific protein-protein interactions [65], which might well be employed to create new biocomputation components [66]. Beyond affinity and specificity, other aspects of molecular interactions might also be worthy design targets. For instance, the engineering of cooperativity combined with low-affinity binding has been proven to generate regulation circuits with a small footprint on the host physiology [67].

5. The cost of bio-computation and its evolutionary consequences

An aspect where biocomputation excels compared with electronic computation is its efficiency. According to theoretical estimates, cells are much closer to the minimal energy required to process information than electronic computers [68]. However, the advantages of this remarkable efficiency gain are somewhat diminished if we consider living organisms’ tight energy budget, because the resources employed in biocomputation (ribosomes, RNA polymerases, precursors, etc) are unavailable for activities that impact on the organism’s reproductive fitness. Moreover, biocomputation might be associated with emergent detrimental effects due to unpredicted toxic interactions between the endogenous organism machinery and the exogenous circuit components [46]. The result is a misalignment between our biocomputational goals and the biological goals of the organism that can lead to the loss of biocomputation circuits in a number of ways (plasmid loss, mutation, transposition, competition, etc) [69].

Addressing this misalignment is key to ensuring the stability of biocomputational devices. There are two main strategies to preserve synthetic circuits [69]: reducing the mutability of the cell, and decreasing the relative fitness cost associated with circuits. In the first case, modelling circuits under the possibility of mutations and gene/-plasmid loss can lead to the design of more robust and stable circuits [70, 71]. The introduction of stabilizing processes, such as a substantial horizontal gene transfer, might also provide an opportunity to increase genetic stability [72,73]. In the second case, some strategies are the use of low-copy number plasmids [45], the expression of circuits in pulses through incoherent feed-forward loops (iFFL) [74–76], or the division of labour in multicellular devices [77–79] — among many others. Finally, there are attempts to combine both strategies simultaneously. The integrase-mediated split of a population of cells into progenitor cells, capable of reproducing without expression nor burden, and a population of producer cells, expressing the circuit of interest but not reproducing, successfully extended the stability of the circuit over time by both unburdening replicating cells, and not replicating burdened ones [80].

More generally, instability and evolution are features that makes biocomputation unique compared to other computation paradigms. The impermanence of synthetic biological circuits should not just be considered as an issue but also as a opportunity [81–83]. Approaches such as directed evolution can exploit this unique feature of biocomputation, for example to optimize and diversify an IMPLY circuit [84], select transcription factor and promoter sets to evolve activators and inhibitors [85], and improve a lysis-circuit performance in the face of variability in the environmental conditions [86].

6. Scaling biocomputations

Back in the 1950s, the first electronic computers were far less powerful than they are today, so the problems they could solve were relatively small [87]. However, advances in chip manufacturing over time allowed those limited machines to become the streamlined, sophisticated computers we are familiar with today [88]. Scalability has been looked at in the field of DNA-based computing, revealing challenges (e.g. cost, spatial diffusion, speed) that seem to prevent DNA-computing from outcompeting electronic computers [89,90]. However, we wonder how the scaling of biocomputation in living cells would look when compared to the scaling of electronic computing.

A key element in the scaling of electronic computing during the last decades was the miniaturization of electronic circuits [88,91]. Many of the circuits implemented in the first computers were pretty big. Over time, manufacturers were able to create new design and fabrication techniques that would decrease the size of those components (Fig. 3A), and today transistors are manufactured at nanometre scales. Attempting to implement a similar approach in biology directly might fail: if we look at genetic circuits, we realize that their components are already at the molecular scale, which leaves little room for further miniaturization.

Fig. 3. Scaling principles in conventional vs. cellular computers.

A. Conventional machines enhance computing capabilities by leveraging space through miniaturisation and repetition. Sample pictures: ENIAC (top) and a microprocessor (bottom). B. In living systems, these principles do not deliver the same magnitude of improvement; but instead, computing abilities can be increased by harnessing the intrinsic complexity of connectivity within networks.

But the miniaturization of components was accompanied also by multiplication, and modern processors can have thousands of nanometric circuits. In the same way, we could consider large biocomputational devices composed of many smaller circuits. Here we might be able to “scale out” computing capabilities of biological systems without “scaling up” the capabilities of their individual components, by leveraging the intrinsic complex behaviour and connectivity between components — something that we do not see in conventional computers (Fig. 3B). Due to the limits on the circuit complexities that can be introduced in a single cell (see Section 5), a potentially powerful strategy consists on distributing complex biocomputation among many cells [79, 92,93]. This approach is analogous to the strategy employed successfully in adjacent fields such as the biotechnological production of chemicals through complex pathways [94]. However, the distribution of biocomputing labour is still limited by the availability of extra-cellular channels enabling cell communication. Most work has focused on the usage of chemical messengers [95,96] capable of diffusing away from the cell and being received by another cell. Though there is an enormous interest in this field, only a few examples, most of them derived from the study of quorum-sensing, are widely used (primarily homoserine-lactone derivates). Nevertheless, there are notable alternatives devised to overcome these limitations, such as the engineering of spatial organization in biocomputation devices [52,97], or the usage of plasmid-conjugation to enable communication between cells [98,99].

7. No free lunch for evolution

A single conventional computer is capable of solving a huge number of different computational problems, if given the right inputs. They are “general-purpose” computing machines [100]. The generality of conventional computers is not an accident. Turing described the Universal Turing machine, a single machine that was capable of any computation, that served as a rough target for the design of our general-purpose computers [100]. In contrast, living cells (and reconfigurable circuits [101,102]) seem to be built for solving a relatively narrow set of specific computational problems, implemented by connecting different types of specialised components or systems.

Living cells do not know about the Universal Turing machine, so it might not be so surprising that they are not general-purpose computing machines. Indeed, there might in fact be a number of reasons to expect evolution to produce a diversity of specialised types of biocomputer rather than a single generalised one. Evolutionary biology and ecology typically explains the biodiversity and functional diversity of organisms using arguments based on environmental heterogeneity and stability of diverse populations [103]. But there is also purely mathematical arguments which may help to motivate a tendency toward specialised over general biocomputation. The “no free lunch theorems” (NFLTs) [104], when applied to optimisation algorithms, say that any algorithm which shows better than average performance for some subset of problems, must pay for this with worse than average performance over the remaining problems [105].

This tells us that one optimisation algorithm can only outperform another by specialising on a certain problem, or type of problem. A very rough application of this to living cells suggests that (assuming they are using their biocomputing capabilities to optimise their fitness) it will always be possible for a specialist to beat a generalist in any given niche, and that the only way for a generalist to increase its fitness on average is to evolve toward specialisation. If this argument holds up, then it highlights major differences between the intelligent design that led to today’s silicon computers, and the evolutionary process that led to today’s biocomputers.⁴

What does this mean for intelligently designed biocomputers? One thing it tells us is that the naturally occurring biological parts that we use to build synthetic biocomputing systems might well have evolved as specialised components to perform a small, specific set of computations, and that reusing them to solve a broader or more general set of problems will necessarily degrade performance. One example of this is genetic promoters. Promoters are genetic components that work well in the natural genetic networks and context in which they evolved, but present significant problems and have been shown to behave very differently when repurposed for synthetic genetic networks [25]. One solution is to tune the promoters to achieve the desired performance in its new setting [106,107]. But of course, since there is no free lunch, any strategy for tuning promoters cannot be the optimal strategy in all contexts and for all measures of performance. Furthermore, if we can frame the promoter’s biological task as an optimisation one, for example, to optimise levels of gene expression with respect to some external environmental cue, the same rule should apply to the promoter itself, in that improving the promoter’s performance in one context will necessarily sacrifice performance in other contexts.

Although there are roles for modularity [108] and reusability [109] in biology, it should be clear that biocomputation is not as modular or general-purpose as our own conventional computation. But we might therefore be able to build more powerful biocomputers by leaning into specialisation in the design process. This could happen at the level of individual components, for example by discovering or engineering large numbers of new components with different functions and parameters [56,57], or new types of component that are currently overlooked in designing biocomputers [110,111]. De-novo generation of specialised biological components is also an exciting new avenue leveraging recent advances in machine learning and artificial intelligence such as diffusion models [112].

At a higher level, choices about which computational models or paradigms to employ can also specialise biocomputations to one subset of problems or another [113]. Certain computational problems, for example multiplying numbers, have simple or efficient solutions within some paradigms, while being significantly more complicated and costly in others [114]. There has been increasing interest in more “exotic” computational models, such as biological perceptrons [115–117], analog [118], neuromorphic [119], probabilistic and stochastic computing [120]. These paradigms, as well as their hybridisation into larger biocomputing systems, have the potential to give biocomputing access to problem domains that might not fit well with the currently prevalent digital logic paradigm.

Evolution has done a good job in the search for complex biocomputers, but it has also inspired a wide variety of algorithms and strategies used in conventional computing to solve optimisation tasks [121]. While all of these different algorithms use the same fundamental ideas borrowed from natural evolution, it is interesting to note that they each perform dramatically differently, depending on the problem to be solved. In light of the NFLTs discussed above this makes perfect sense, but it also means that the same should apply to natural evolution itself. Natural evolution has done a good job in the search for complex bio-computers, but if we believe the NLFTs, it cannot be the best strategy to use for all problems. That is, as well as learning from evolution and mimicking evolutionary strategies in the design of our biocomputations, there is also the opportunity for us to do better than natural evolution in certain (but perhaps narrow) domains.

8. Summary and conclusions

Conventional computers are capable of solving all computational problems that it is possible to solve. Since they are such powerful machines, it is natural that synthetic biology has tried applying approaches and ideas borrowed from the design of conventional computers to the design of biocomputers. The large repertoire of experimental possibilities enabled by biology has given rise to different approaches to implementing biocomputers: from computation based on the molecular biology of DNA, to cell-free systems, to the engineering of living cells. However, it is obvious that this has not yet produced biocomputers with computational power on par with conventional ones, and we argue that it is unlikely to do so in the near future. In this work, we target computing in living cells, given our focus on questions about “design principles” that are in some manner expressed through natural evolution. For readers interested in similar questions applied to biocomputing in other settings, previous work covers the usage of DNA-based computing [89,90], network-based computing [55,122], or cell-free systems [123]—in some cases, producing analytical models of the scaling capabilities of some of those fields of biocomputation [90,122, 124].

For sure, there are some commonalities between conventional and biological computers. In principle, both are capable of solving exactly the same computational problems, namely all those that are solvable. Computational complexity still applies to problems regardless of the kind of computer that solves them. But as we struggle to design more powerful biocomputers, it might be more instructive to look for differences, and the reasons for them. Our computers are simple (in the sense meant in Section 2), modular, static, made of only a few types of discrete component, and built to perform a huge number of general computations. Biology’s computers are complex, dynamically changing, interconnected systems of many different types of optimised component, built to perform a much smaller number of computational tasks very well.

The nature of information, and how it flows to facilitate computing is another fundamental difference. Some of this can be simply and reasonably attributed to physical limitations and the differences in the substrate. However, there are still questions the role of information theory in shaping biological processes. Some flows of biological information seem to be optimised with respect to information theoretic measures, while others are not. Whatever the answer, discovering why biology seems to deviate from the theoretically optimal solutions could help us on the way to engineering them.

Most theories about the origin of life and its early evolution point to very simple chemical systems that have abilities to self-replicate and evolve [125–127]. It is unlikely that those systems already had the extent of genetic organization that we see in modern organisms. Complexity was probably acquired over time, and actually during a very short period of time, considering that many reports claim the last universal common ancestor (LUCA) [128] had comparable complexity to that we see nowadays. The same progression in computing power has shaped the history of computing machines, from the differential engine (and even earlier) to today’s silicon-based computers (Fig. 4A). Our own efforts at biocomputation lie at the intersection of both.

Fig. 4. Computing at the intersection of design and evolution.

A. The history of evolution, from LUCA (last universal common ancestor) to modern organisms, has progressively enhanced the computing abilities of living systems over time (top), much like the history of computing machines, from Babbage’s Difference Engine to today’s computers (bottom). B. Biocomputation unites these two timelines by designing computing devices using living matter in a rational and purposeful manner.

If we look at life now, we find an abundance of extremely sophisticated biocomputers. However, it’s also very likely that early life had to adapt to the environment and perform certain computations, at least at a fundamental level. Therefore, it would be very interesting to understand how early life performed those computations and how it, over time, acquired the possibility of solving more complex problems. From a synthetic biology perspective, understanding how life scales its biocomputational abilities will teach us a lot about creating new biocomputing techniques. Ground-breaking advances in the scalability of biocomputing might help us understand the puzzle of how life originates from a very simple chemical basis. This in in fact a major contributing factor to the differences between conventional computers and biocomputers. Conventional computers were intelligently designed, whereas biocomputers have evolved (Fig. 4B).

Natural evolution is responsible for the naturally occurring biocomputers we have today. Since it seems to have been to successful, it is not too strange to suggest that instead of borrowing from the design ideas of conventional computing, we should instead borrow them from natural evolution. We fully support this point of view, but also point out that there could be subtle ways in which the evolutionary process shapes and limits the kinds of biocomputers it produces. For example, although far from a formal proof, in Section 7 we outline a rough argument for why natural evolution has a tendency to produce specialised biocomputers rather than general ones, based on the NFLTs.

Natural evolution can itself be considered a computation, one which we could perhaps engineer and exploit ourselves in the future. We also argue in Section 7 that this leaves open the exciting possibility for improving on the performance of natural evolution in the design of biocomputers. The most promising and direct strategy for achieving this would likely be to start with exploiting natural evolution, and tuning its parameters (whatever those turn out to be), to fit the specialised structure of a narrow problem domain. This of course requires serious effort toward understanding and engineering the mechanisms of natural evolution for inclusion in the synthetic biology toolkit.

Finally, we would like to highlight that engineered biocomputations do not necessarily need to scale up to meet the capabilities of conventional computing [89]. We do not breed horses to compete with cars, and engineered living cells do not need to compete with laptops for niches. Therefore, while a never-ending exponential scaling in biocomputational capabilities might be desirable, improving just a few orders of magnitude might be enough to open up new possibilities, and exciting new applications.