Synthetic gene circuit evolution: Insights and opportunities at the mid-scale

Abstract

Directed evolution focuses on optimizing single genetic components for predefined engineering goals by artificial mutagenesis and selection. In contrast, experimental evolution studies the adaptation of entire genomes in serially propagated cell populations, to provide an experimental basis for evolutionary theory. There is a relatively unexplored gap at the middle ground between these two techniques, to evolve in vivo entire synthetic gene circuits with complex dynamic function instead of single parts or whole genomes. We discuss the requirements for such mid-scale evolution, with hypothetical examples for evolving synthetic gene circuits by appropriate selection and targeted shuffling of a seed set of genetic components. Implementing similar methods should aid the rapid generation, functionalization, and optimization of synthetic gene circuits in various organisms and environments, accelerating both the development of biomedical and technological applications and the understanding of principles guiding regulatory network evolution.

Graphical Abstract

Helenek, Krzysztoń et al. describe “mid-scale evolution”, the process of developing and optimizing entire synthetic gene circuits as they evolve naturally, with potential human interventions inside living cells. This offers the capability to rapidly generate, characterize, and optimize gene circuits in various applications and conditions.

1. The middle ground between experimental and directed evolution

Synthetic biology is a relatively new field focused on designing biological systems for predefined purposes^1,2. Conceptualized decades ago^3,4, synthetic genetic systems have become a reality only in the late 20^th century, due in part to the advancement of recombinant DNA technologies⁵. Most efforts are directed towards altering, rearranging, assembling, and remaking genetic material to create cellular functions. Examples include cells that switch phenotypes^6,7, oscillate^6,8, respond in predefined ways^9,10, form spatial and temporal patterns^8,11,12, produce metabolites^13,14, sense^15,16, and destroy¹⁷ or support¹⁸ other cells. As with all of biology, synthetic biology is applicable on multiple scales. At the smallest scale, synthetic biology (re)designs single molecules¹⁹, at the mid-scale, multi-molecular regulatory systems (such as gene circuits, signaling and metabolic pathways)^7,20; while at the largest scale, entire genomes²¹. The field has relied primarily on engineering approaches, which have limitations when applied to biology, where uncharacterized and noisy components, and complex factors affecting system design are common²². To overcome these limitations, evolutionary methods have also gained traction over the years^23,24 as efficient ways to rapidly optimize function at various levels, from small genetic components to entire genomes. The development of such evolutionary optimization called “directed evolution” and its “phage display” precursor were recognized by the 2018 Nobel Prize in Chemistry.

Directed evolution leverages Darwinian principles, creating artificial genetic diversity and selection pressure to enhance synthetic gene circuit function for therapeutic and industrial applications²³. The process to create variant libraries of key DNA regions can be accelerated by repeated rounds of focused, small-scale genetic diversification²⁵ via in vitro point mutagenesis, or by in vivo high-throughput methods like PACE²⁶, VEGAS²⁷, and OrthoRep²⁸, MutaT7²⁹ and EvolvR³⁰, in organisms ranging from bacteria to mammalian cells³¹. Apart from point mutations, larger-scale variation by shuffling of DNA pieces or imposing other structural variations underlie directed evolution by combinatorial optimization³² of bacterial operons³³, as well as structural characterization of promoters³⁴, or synthetic gene circuits³⁵. The technique’s success hinges on generating genetic variation and on evaluating the desired properties, to enrich beneficial variants by high-throughput, sensitive, and robust methods. During its short history³⁶, directed evolution has enabled significant improvements in enzymes^37,38, RNAs³⁹, pathways^33,40, or genomes⁴¹ for various industrial and medical applications.

Despite its major contributions to the progress of synthetic biology, the power of directed evolution has been underutilized, in at least two ways. First, directed evolution has not been used sufficiently to advance evolutionary biology, e.g., to verify emerging principles of evolutionary theory⁴², or to address fundamental questions of adaptation. Second, directed evolution has focused on improving the performance of either individual genetic parts (enzymes, protein domains, sequences related to transcription and translation), or their configurations, rather than on evolving in vivo entire systems with complex regulatory function, especially from scratch. Third, directed evolution tends to ignore the genetic context, which can directly and nontrivially affect element optimization, limiting the utility of the approach^34,35,43,44. Despite a significant early paper being titled “Directed evolution of a genetic circuit”, that study optimized individual components of the circuit rather than the entire circuit at once, as a whole⁴⁵. Overall, examples are lacking for directed evolution of whole mid-scale systems with sophisticated dynamic behaviors, such as oscillatory gene circuits, switches, fine-tuners, and pattern generators.

In contrast to the practical motivation of directed evolution, experimental evolution aims to elucidate fundamental evolutionary mechanisms of various organisms adapting to laboratory conditions⁴⁶. The Long-Term Evolution Experiment (LTEE) ongoing since 1988 in the Lenski, and then, Barrick labs⁴⁷, has greatly advanced the testing and discovery of evolutionary principles in laboratory conditions, addressing questions of predictability, epistasis, facility, and speed of adaptation. For example, Lenski and collaborators utilized 12 replicate E. coli populations derived from the same ancestor to investigate adaptation paths in defined growth medium. They found that fitness and cell size of all populations increased initially. Then, cell size stopped changing after 2,000 generations, whereas fitness kept creeping up throughout the LTEE, although at a decreasing rate, suggesting “diminishing returns” from new beneficial mutations⁴⁸. However, the mechanisms of fitness increase were often difficult or impossible to unravel. Similar evolution experiments have been performed with yeast^49–51 and other living organisms^52,53 or bacteriophages⁵⁴ as well as “digital” organisms, which are self-replicating and evolving computer programs^55–57. High-throughput methods such as eVOLVER⁵⁸ can scale up evolution experiments. Yet, overall, the lack of constraints in experimental evolution compromises the interpretability of its outcomes, creating a need to evolve smaller-scale systems rather than entire genomes.

Whereas evolutionary experiments have been primarily large-scale studies of whole genome evolution in general laboratory conditions, there were occasional examples of special conditions applied to improve or study small natural networks or synthetic gene circuits as they evolve, approaching what we call mid-scale evolution. For example, measurements of the external lactose-dependent cost and benefit of Lac proteins in E. coli allowed the prediction of an optimal Lac expression, which can differ from the wild type Lac level⁵⁹. During subsequent propagation of cells in various constant lactose concentrations, Lac levels evolved to their predicted optima. Unfortunately, it is unknown if the optimizing mutations were inside or outside the Lac operon. A different study examined the evolution of the Lac system in constant or alternating sugar conditions, observing frequent mutations in the Lac repressor and its DNA binding region⁶⁰. In another example approaching mid-scale evolution, an inducible synthetic genetic system in E. coli⁶¹ controlled the expression of two genes simultaneously: one that was detrimental in the presence of sugar, and one that was beneficial in the presence of an antibiotic. Thus, high circuit expression in antibiotic and low circuit expression in sugar was favorable. Reversing circuit induction such that the beneficial gene had low expression, while the costly gene had high expression in corresponding environments created selection for the LacI repressor to reverse its function. Upon repeated rounds of mutagenesis and environmental shifts, the fitness conferred by the circuit increased, and LacI’s inducer response reversed (i.e. inducer-bound LacI caused repression)⁶¹. Implementing a similar methodology in yeast cells⁶² converted a DAPG-OFF system into a DAPG-ON system. Despite their successes, these studies either mutagenized and screened only one component to improve gene circuit performance or evolved whole natural genomes, and therefore still do not classify as mid-scale evolution.

Combining synthetic biology and experimental evolution at the mid-scale (Figure 1, Table 1) to evolve whole synthetic gene circuits rather than individual genetic components or whole genomes is important for witnessing, understanding, and utilizing evolution. Mid-scale evolution under appropriate selection may be crucial not only to deduce general evolutionary principles, but also to design and optimize the function and adaptation of regulatory networks, or prevent their breakdown^63,64 in living organisms.

Figure 1. Mid-scale evolution lies between directed and experimental evolution.

There is a gap between directed evolution and experimental evolution, which suggests that mid-scale evolution of whole synthetic gene circuits inside living cells would be both useful and informative. See also Table 1.

Table 1.

Mid-scale evolution is between experimental and directed evolution.

Criteria	Experimental / genome evolution	Mid-scale / gene circuit evolution	Directed / component evolution
Predictability	Unpredictable	Somewhat predictable	Mostly predictable
Target of evolution	Whole viral or cell genomes evolve	Entire gene circuits evolve, coupled with genome	Either circuit components or their arrangements evolve, genome irrelevant
Field	Evolutionary biology	Evolutionary, synthetic, systems biology	Bioengineering, synthetic biology
Type of Genetic Alterations	Natural genetic variation of any type in vivo	Natural and/or artificial point mutations and structural variation mainly in vivo	Either point mutagenesis of part(s) or arrangements of parts, mostly in vitro
Purpose	None - fundamental biology	Fundamental biology and/or improvement of entire circuits	Purpose driven-improvement of parts or their arrangements
Modeling Predictions	Evolvability, robustness, emergence of complex features, types of mutations and speed of fixation	Network-level mechanisms of adaptation, types of mutations and speed of fixation	Molecular mechanisms and mutational paths to improved component performance

Between the extremes of experimental and directed evolution (Table 1) some studies have explored the continuous mid-scale evolution of small multi-gene synthetic circuits as they lose⁶³, alter⁶⁵, or optimize⁶⁶ complex, dynamic functions in vivo. For example, a positive feedback-based bistable synthetic gene circuit in yeast⁶⁷ was evolved to either alter or lose its bistable behavior in six different environments⁶⁵. Environments with various inducer (i) and drug (d) combinations targeted specific costs and benefits of auto-activated gene expression, which was costly in inducer (i,0); beneficial in drug (0,d); and beneficial but too costly in inducer and drug (i,d). Mathematical models allowed mapping the environment-dependent fitness landscapes of the gene circuit⁶⁷, which then predicted the type of mutations observed in each environment⁶⁵. For example, gene expression heterogeneity was altered in (i,d) due to a few promoter and coding mutations of the auto-activator gene, whereas bistability was lost in (i,0) due to many mutations abrogating the function of the auto-activator gene. Remarkably, applying the (i,d) environment to renew selection for high expression on apparently nonfunctional mutants originating from (i,0) revealed various ways of losing bistability. In some cases, the circuit was evolutionarily repaired, regaining bistability due to additional evolution under renewed selection⁶⁶.

Another pair of inducible noise-controlling gene circuits, integrated into the genome of mammalian cells and controlling drug resistance gene expression, lost their tunability, gaining constitutively high expression upon continuous selection in various drug concentrations⁵³. Subsequent experiments revealed DNA amplification as the mechanism causing increased expression, suggesting novel nucleotide therapy to combat chemoresistance, which was verified in human cancer cell lines⁶⁸.

How could mid-scale evolution expand to other systems? Next, we outline the possible benefits, the requirements, and a few potential examples of mid-scale evolution.

2. Opportunities at the mid-scale

Despite the successes of forward engineering in synthetic biology, it is highly uncertain if a given circuit developed in well-defined laboratory conditions, in a certain organism will function properly in different conditions or different organisms^69–71. Even if function persists, some properties, such as the period or amplitude of oscillations, the memory of switches, or the width of pulses may not be satisfactory or applicable in new environments. Mid-scale gene circuit evolution in vivo, by continuous diversification through both point mutations and structural variation should facilitate and accelerate achieving or restoring the desired function compared to directed evolution. Furthermore, mid-scale evolution would contribute to understanding the evolution of natural gene regulatory networks⁴², which might be organism- and environment-specific. Therefore, mid-scale evolution is needed to evolve entire synthetic genetic systems with dynamic functions, such as oscillators, pulse generators, switches, and noise controllers. This requires three ingredients as we discuss below.

First, we need design principles guiding the setup of mid-scale evolution. Knowing how existing natural or synthetic systems perform the desired function will facilitate evolving similar functions. Second, we need the ability to select variants approaching the goal, which requires components with adjustable cost and/or benefit in specific environments that we can impose. Specifically, as we will discuss, we need some circuit components that are beneficial in some conditions but become toxic otherwise. Costly stress-mitigating proteins behave like this with and without stress, respectively, and will be a recurrent theme below. Third, we need selectable variation during mid-scale evolution. Variation can stem from naturally occurring mutations. However, in some cases this may be too slow or insufficient, creating a need for targeted evolutionary accelerator systems. Assume that we have evolving cells that can randomly and continuously shuffle or amplify/delete a seed set of genetic components, while creating point mutations inside each component. Point mutations in these components will alter biophysical parameters, such as binding constants (to both inducer molecules and DNA operator sites), strengths or thresholds of activity, and degradation times. Additionally, shuffling can generate different regulatory relationships between the components. Thus, the system will allow variation both by random circuit rewiring and by individual component modification. We postulate that such an initial “seed” of components will wire up and evolve into circuits with desired dynamics in host cells under cleverly chosen selection. Next, we discuss a few potential examples of mid-scale evolution with such premises.

2.1. Evolve an oscillator.

Rhythms appear to be critical in numerous biological functions, fulfilling roles and providing benefits that are still being investigated⁷². Repeated cycles of high and low levels of regulator molecules (i.e., protein, miRNA) drive the cell cycle^73–76, circadian rhythms^75,77, DNA damage response^78–84, immune response^85–88 and cell differentiation⁸⁹. Although diverse molecular networks underlie various biological oscillators^90,91, three overarching design principles have emerged for oscillators^{20,76,92–94}: the presence of time-delayed negative feedback; positive feedback for more robust and persistent oscillations; and proper production and degradation rates to sustain oscillations.

Evolving an oscillator thus requires a seed of components with a potential to fulfill these three design principles. For instance, a set of repressor and activator transcription factors and a corresponding set of regulatory binding sites that ensure nonlinear responses^95–98 would be sufficient to enable both negative and positive feedback and thus, oscillator evolution. Delay could arise from regulatory cascades, separation into different cell compartments, or by scaffold molecule action^99,100. To head start evolution, in some cases the seed component set could already include either a form of negative feedback with appropriate delay, or positive feedback, which should be robust and persist through evolution, to avoid losing the advantage they confer.

To select for oscillator evolution, we must impose recurring conditions resembling the periodically returning benefits of light that most likely caused the evolution of circadian oscillators¹⁰¹. Therefore, some component should be periodically beneficial over time, in some periodically returning environmental factor, but costly otherwise (Figure 2a). Possible examples include galactokinase in periodic galactose pulses¹⁰², a drug resistance gene in periodic drug pulses, or even a fluorescent reporter subject to periodic flow-sorting^103,104. In a different scenario, a beneficial component could subsequently burden the cell, as in the natural p53-Mdm2 system^78–80,105, where upon DNA damage, high p53 expression promotes DNA repair, but also mounts cell cycle arrest^81–84. Under constant stress, periodically cycling levels of such components with always suboptimal expression reap the benefits while mitigating the costs. To avoid the evolution of periodic multicellularity¹⁰⁴ and resolve temporal signals from single cells, evolving cells should be evaluated by time-lapse microscopy or other single cell tracking methods. Increasing peaks in Fourier-transformed or wavelet-transformed signals would indicate the emergence of periodicity.

Figure 2. Examples of mid-scale evolution for various known gene circuits. a. Oscillator.

Selection should consist of recurring conditions ensuring that some component is periodically beneficial over time, but costly otherwise. Such periodic selection could involve a costly gene within the circuit, which is growth-promoting in some periodically returning environmental factor.

b. Switch. Selection should consist of a transient inducer pulse dropping to an intermediate level, which activates a stress resistance gene, followed by persistent stress to select for persistent activation. Evolving toggle switches would require two similar stimuli and two corresponding stresses.

c. Pulse generator. Selection should consist of a pulse of stress concurrent with an increase in inducer that subsequently stays on. Highly costly stress-response gene expression will give incentive for subsequently diminished expression. The stress pulse will last a limited time, while the inducer will persist to select against sense-response or switch circuit behaviors, where the costly product would be continuously high, compromising cellular fitness. Therefore, the cells will be pressured to produce a short pulse of the costly product to mitigate the stress in the environment, but not maintain its expression when the stress is gone.

d. Sense-response system. Selection should consist of concurrent increase of both inducer and stress, followed by concurrent decrease of both after some plateau. These conditions should be repeated nonperiodically, with various stress levels versus inducer levels corresponding to the desired sense-response function (e.g., gradual or stepwise).

e. Noise generator. Selection should consist of a concave fitness landscape to diminish gene expression noise, and convex fitness landscape to amplify noise. Alternating selection for high and low expression could substitute for convex landscapes.

2.2. Evolve a switch.

Switches implement a sharp, stable response to an environmental factor, even if its presence is transient. Although by nomenclature genetic switches include transcriptional switches, riboswitches, and toehold switches¹⁰⁶, here we only consider bistable switches that flip and maintain their state even after the external stimulus is diminished. This cellular “memory” is important in the context of personalized medicine, for example in CAR-T cell therapy¹⁰⁷ The simplest switches are thus autoactivating genes^67,108,109 and toggle switches^7,110. Seeding the evolution of switches requires either transcriptional activators and their cognate promoter sites for gene autoactivation, or transcriptional repressors and their promoter sites for toggle switches⁷. To enable switch evolution, a transient inducer pulse dropping to an intermediate level should turn on a stress resistance gene, followed by persistent stress to select for persistent activation (Figure 2b). Evolving toggle switches would require two stimuli and two stresses in a similar manner. For bacteria, the stimuli could be two sugars (such as lactose vs. glucose) acting on a seed of sugar-binding repressors and operator sequences that also control two costly components¹¹¹. Sorting for high-expressing cells with some delay after a transient stimulus and monitoring the stability of high expression should reveal if the system is starting to work as a switch. There must be a way to turn the switch off, to reinitiate the process.

These ideas can apply to more complicated switch configurations, such as cascading bistable switches that interact with each other¹¹² and drive epithelial-mesenchymal transition in natural contexts^113,114. Seeding cascading switches should ensure the orthogonality of components and their selection, with the details dependent on the desired phenotype. Specifically, selecting a single switch will favor a “winner-take-all” phenotype, while simultaneous selection will favor a doubly-expressed phenotype¹¹². Overall, with the appropriate selection strategy, the evolution of various switches should be feasible.

2.3. Evolve a pulse generator.

Pulse generators are genetic systems that produce transient gene activation upon the onset of a relatively constant stimulus¹¹⁵. In contrast to transcriptional bursting, which is the stochastic fluctuation of gene products^116,117, pulses tend to have predictable durations and amplitudes, being produced by particular network motifs¹¹⁸. Pulsing is seen in a variety of cellular contexts, ranging from bacterial sporulation¹¹⁹, to hormone release¹²⁰, stress responses¹²¹, and cell growth^122–124. Synthetic pulse generator circuits^115,125 can be utilized as counters, to mark how many times a biological reaction occurs in the cell, or as discriminators between transient and sustained oscillatory signals¹²⁶. Pulsing has similarities to oscillations, which can make it difficult to differentiate between the two¹²⁷.

A design principle for pulsing is the convergence of activator and repressor action, with a sufficient time delay, on a common genetic target^125,127, as in incoherent feed-forward loops (iFFLs)^125,126,128. From analyzing iFFLs, important parameters for generating pulsing behavior are activation thresholds, degradation rates, and the time delay between activation and repression¹²⁶.

Therefore, for the evolution of a pulse generator, we can seed activators and repressors with their cognate binding sites. The target of selection can again be a gene product that is beneficial in stress, but costly otherwise. Selection could consist of a stepwise rise of the inducer, which turns on the expression of a stress-response gene that protects from a pulse of stress (Figure 2c). This strategy will select for a transient expression profile in response to a long-lasting signal, which is the hallmark of pulse generators¹¹⁵. Evolving cells could be screened as for the oscillator.

The amplitude and duration of pulse characterize a pulse generator. Adjusting each of these features could be useful in different contexts. For example, modulating the strength of stress could tune the amplitude of the pulse, while the duration of stress could modulate the duration of the pulse. Therefore, by adjusting the experimental setup, different parameters of the pulse generator can be evolved as desired.

2.4. Evolve a sense-response system with desired properties.

Widely prevalent in biology, sense-response systems can detect internal or external stimuli, subsequently eliciting a specific response¹²⁹, which could be sharp or gradual. Typically, sense-response systems encompass a broad range of biological components, including receptors, signaling and effector proteins, and RNA in riboswitches¹³⁰, immune responses¹³¹, metabolic homeostasis¹³², and environmental adaptation¹³³. Accordingly, a variety of synthetic sense-response systems have been designed and engineered to create novel biological responses¹³⁴, circumvent dysfunctional native networks, or restore proper cell homeostasis, paving the path to cell-based precise therapeutics^135,136.

Generally, sense-response systems require a sensory component with stimulus-dependent activity, driving a regulated component for response generation. This response can be either activated or suppressed by the stimulus.

Among the simplest sense-response systems are riboswitches¹³⁰ and autoregulatory transcription factors¹³⁷, where the same component senses and responds to the stimulus. The general design features of a sense-response system depend on the desired response characteristics, such as the basal and saturating responses, the response range, and the response time, which have been studied extensively^69,137,138. Mid-scale evolution will enable tailoring these response characteristics in new biological and environmental contexts.

Evolving sense-response systems can begin with the seeding of a stimulus-binding transcription factor and regulators with their cognate promoters. Strategically designed selective pressures can drive the directed evolution of circuit characteristics according to the general fitness burdens and stress-elicited benefits of circuit components (Figure 2d). Typically, the fitness landscape should correlate with the response function. For instance, a gradual response would require gradually increasing benefits, whereas a sharp, switch-like response would require threshold-like benefits onset versus increasing inducer concentrations. Connecting a growth factor with the responder component might lead to increased basal and saturating responses. Furthermore, the timing of selective pressure can modulate the response time of the system. Screening could consist of repeatedly measuring how far the evolving stimulus-response characteristics are from the desired goal, for example, by functional distance metrics. Beyond the response characteristics of the mean, it may be important to characterize and optimize the variability of responses across the cell population, as we discuss below.

2.5. Evolve a noise generator/processor.

Gene expression stochasticity or noise refers to nongenetic, unpredictable differences present even in isogenic cell populations, which may originate from intrinsic randomness in molecular interactions or extrinsic cellular differences or perturbations^139,140. However, noise can have beneficial or adverse effects depending on the environment^{53,117,141–143}, or the downstream biological effect¹⁴⁴, such as the fate of virus-infected cells¹⁴⁵ or stem cell differentiation¹⁴⁶. Being able to control gene-specific noise levels in a system is, therefore, useful to enable biomedical applications, as well as to study how noise adapts or affects adaptation^53,147,148.

Practically, all genes have somewhat noisy expression, and various characteristics of gene expression such as promoter kinetics, RNA and protein synthesis and degradation rates, as well as regulatory cascades, feedforward loops, positive and negative feedback loops can adjust the level of noise^149–151. Accordingly, leveraging multiple genetic features, various synthetic biological systems achieved noise-mean decoupling, i.e., the independent control of the noise and the mean^{53,117,152–154}.

Examples of mid-scale evolution of noisy gene expression systems suggest how to design noise evolution^{65,66,103,138}, which indicate that noise evolution should be easy if appropriate fitness landscapes can be imposed^138,155,156. Specifically, long-term evolution on concave fitness landscapes (near fitness peaks) should diminish gene expression noise, whereas evolution on convex fitness landscapes (in fitness valleys) should amplify noise (Figure 2e). Alternating selection for high and low expression could substitute convex fitness landscapes¹⁰³. To avoid the concurrent evolution of the mean, the landscapes should have no overall slope¹³⁸.

Seeding the evolution of noise controllers may depend on the desired type of gene circuit. The seeding component set could have a potential to assemble into positive or negative feedback to enable noise amplification or suppression, respectively. In other cases, regulatory cascades, promoter features or RNA interactions could modulate noise^157,158.

To screen for the evolution of noise, typical measures such as the CV or Fano factor could be investigated by flow cytometry in clonal populations, whereas studying temporal features will require time-lapse microscopy¹⁵⁹ to estimate memory, and Fourier spectra to screen against oscillations.

3. How to get there – Requirements for mid-scale synthetic evolution

Enabling the mid-scale evolution scenarios discussed above requires the in vivo capability to randomly shuffle or amplify/delete a set of genetic components, besides creating point mutations inside each of them. The seed should evolve in living cells, relatively independently of the host genome. Designing the selection pressure to target the seed of components, and then the evolving gene circuit later, will minimize the chance of alternative solutions by background genomic adaptation. Multiple recent studies of gene circuit evolution indicate that this is possible^53,65,66,68. Additionally, steering evolution towards some desired gene circuit dynamics will require automatically monitoring single cells and selecting/transferring them according to how well they fulfill a predetermined goal function¹⁶⁰. Finally, selection will require gene products beneficial in specific environments but costly otherwise. Next, we will highlight potential ways to achieve these criteria and how current technology is relevant to them.

Controlling genetic variation and selection will be necessary to optimize mid-scale evolution. First, gene circuit rewiring by structural alterations (i.e. amplifications, insertions, deletions, relocations, and inversions of larger DNA segments) should occur at time scales allowing the exploration and screening of numerous circuit topologies. Second, specific locations within the seed should be preferentially targetable for genetic variation, besides naturally occurring mutations. Third, the seed of components should not be entirely deactivated or removed from the cell population. Neither directed nor experimental evolution has fulfilled these requirements so far, but recent theoretical work suggests that adequately controlled external factors can accelerate evolution¹⁶¹, and experimental opportunities are arising with diligent design and tuning.

For instance, besides choosing appropriate components and corresponding selection pressure^53,65,66,68, host-orthogonal transcription and translation offer additional ways of isolation from the host genome^162,163. Among others, the OrthoRep system consists of transgenic chromosomes autonomously replicating with a 1000x higher point mutation rate¹⁶⁴, in physical isolation from the host. Although promising, OrthoRep offers limited control on the location of hypermutations, risking rendering the seed non-functional. Currently, bacterial and mammalian equivalents of OrthoRep are lacking.

A more feasible goal currently is to integrate the seed into an appropriate genetic locus (such as a genomic “safe harbor site”¹⁶⁵ in mammalian cells), and locally maintain expression and mutation rate. Recently, we developed a post-CRISPR, recombinase-based strategy for robust, repeatable, site-specific integration of genetic circuits into the human genome¹³⁸, without any potentially harmful DNA breaks. Yet, recombinase action allows much more than exchanging single DNA fragments. Appropriately positioning recombinase recognition sites allows DNA insertions, excisions, inversions, and translocations^166,167, which can effectively shuffle, displace, amplify or remove any functional DNA sequence (i.e. promoters, binding sites, open reading frames) within the circuit with recombinase-dependent rates (Figure 3). Thus, it becomes possible to locally increase the rate of selectable structural variation, creating variable circuit topologies that emerge according to seed design and rearrangement types (e.g. larger DNA fragments shuffled less than shorter ones). Shuffling rates might be adjustable by controlled expression¹⁶⁸ of orthogonal recombinase systems (FLP-FRT and CRE-LOX).

Figure 3. A recombination-based system for shuffling genetic components.

a. Nonfunctional, scrambled arrangement of seed components with functional elements separated by recombination sites.

b.,c. The seed of components can rearrange itself into functional gene circuits by random recombination events of inversions (1), insertions (2), cassette exchanges (3), and excisions (4) as recombinases rearrange and insert/excise the integrated functional elements switching them with extrachromosomal circular DNA fragments.

Besides component rearrangements and loss/gain, elevated point mutagenesis will be necessary to alter the kinetic properties of the circuit, requiring controllable point mutation rates at various portions of the seed. This might require dCas9-based¹⁶⁹ or other systems^25,29,30 causing localized, site-specific hypermutation, which has successfully been used in directed evolution designs. Moreover, the shuffling system could fail due to random point mutations within recombination sites, requiring locally lowered mutation rates, which could be accomplished by Cas9-based mutation-prevention systems¹⁷⁰ or gene entanglement¹⁷¹.

In mammalian cells, integration into safe harbor sites¹⁶⁵ should aid isolation by preventing transgene silencing and minimizing unwanted side effects of integration. However, silencing can still ensue if the metabolic burden of circuit expression is too heavy¹⁷². Synthetic read-write systems of chromatin remodeling orthogonal to those native to mammalian cells¹⁷³ could overcome this, stably maintaining transgene expression over 1.5 kilobase long active chromatin regions, to strengthen the utility of safe harbor sites under high selective pressure. Safe harbor sites other than AAVS1 are available, such as CCR5 and Rosa26¹⁶⁵, or others¹⁷⁴.

When combined, the above solutions can enable mid-scale variation. Besides variation, we need appropriate selection. Targeting the selective pressure onto the seed components can isolate circuit evolution from the native genome^53,65,66,68. For example, changing nutrients^175,176, cellular density^177,178 or temperature¹⁷⁹ could mitigate or emphasize the effect of environmental factors that alter gene circuit performance, such as resource competition¹⁸⁰. A common theme above was a set of components that is beneficial in a certain condition, but costly otherwise. Many natural and synthetic⁶⁷ systems satisfy these requirements, such as toxic activators¹⁸¹ driving stress-response genes. Costly expression causes growth feedback^175,180,182, which could help achieve desired dynamics or amplify selection against high expression. Also, cells can evolve to eliminate problematic expression cost and growth feedback while maintaining circuit function⁶⁶. For the alternative oscillator design, synthetic components with beneficial, stress-protective expression causing subsequent toxicity could include transcriptional regulators coexpressed with a stress resistance gene and a toxic gene such as thymidine kinase (TK), which converts a prodrug into a cell cycle arrest factor¹⁸³. Then, increasing gene expression that mitigates stress will convert into cell cycle arrest and apoptosis, incentivizing a drop in gene expression, and so on, giving rise to oscillations. Importantly, the progress of mid-scale evolution needs to be evaluated quantitatively, extracting the most promising cells from the population. Flow cytometry might be sufficient when steady states provide appropriate success metrics. Otherwise, more elaborate methods, such as time-lapse microscopy coupled with microfabricated surfaces^184–186, hydrodynamic traps¹⁸⁷, microwells¹⁸⁸, or droplet microfluidics^189–191 might be needed to characterize and extract best-performing individual cells. For example, cells can be selected from a pool of fluorescently labeled droplets sorted by Fluorescence-Activated Droplet Sorting (FADS)¹⁹², or isolated from microwells or micropatterns using FluidFM technology¹⁹³ or optical tweezers¹⁹⁴. Streamlining single-cell selection requires real-time processing and analysis of high content microscopy data. This is still challenging since it demands high computational power and streaming capacity, but substantial advances are on the horizon^195–197.

Overall, mid-scale evolution of synthetic gene circuits may reveal new aspects of network evolution, while simultaneously enabling rapid gene circuit functionalization and optimization. Theodosius Dobzhansky’s famous sentence “Nothing in biology makes sense except in the light of evolution”¹⁹⁸ is just as valid for synthetic biology as for other areas of biology. As the field progresses in exploring the middle ground between experimental and directed evolution, this concept will gain new meaning as more complex behaviors are probed, uncovered, and understood.