Genome‐scale engineering for systems and synthetic biology

The phrase ‘genome‐scale engineering’ invokes a future in which organisms are custom designed to serve humanity. Yet humans have sculpted the genomes of domesticated plants and animals for generations. Darwin's contemporary William Youatt described selective breeding as ‘that which enables the agriculturalist, not only to modify the character of his flock, but to change it altogether. It is the magician's wand, by means of which he may summon into life whatever form and mold he pleases’ (Youatt, 1837). Selective breeding has transformed aurochs into Holsteins, wolves into Chihuahuas and Great Danes, and teosinte into maize. All of these examples involved genomic changes at a scale dwarfing any attempted through rational design. Understanding why genomes have been more readily shaped by evolutionary principles than conventional design‐based approaches is important for current and future genome engineering endeavors.

Engineering is a human enterprise consisting of iterative cycles of design, construction, and testing. Optimizing this iterative process involves balancing the relative time, costs, and expected benefits gained at each phase. However, rationally designing and building a genome to produce the desired phenotype has proven exceedingly difficult. Designing organisms to specification requires accurately predicting phenotype from genotype, a complex problem that is worsened by our incomplete knowledge of biomolecule production, degradation, and interaction rates. Moreover, the computational resources required to run bottom‐up molecular‐level simulations are daunting even for simpler systems (Karr et al, 2012; Koch, 2012). Nevertheless, models have been useful for generating new hypotheses and targeting promising areas for engineering. Yet, even with the best in silico predictions, we are still limited by our ability to construct the designed genome. More than any other factor, the absence of molecular tools for manipulating genomic sequences has forced us to rely on selective breeding and evolutionary optimization (Conrad et al, 2011) rather than rational genome design.

Recent breakthroughs in genomics and genome editing have promised a greater role for rational design in biological engineering (Figure 1), offering new opportunities for systems and synthetic biologists aiming to reverse‐engineer naturally evolved systems and to build new systems. In particular, advances in high‐throughput DNA sequencing and large‐scale biomolecular modeling of metabolic and signaling networks represent two important new frontiers that aid genome‐scale engineering. Over the last few years, thousands of bacterial genomes have been sequenced from a wide variety of natural species and numerous laboratory‐generated strains (Pagani et al, 2012). These efforts have illuminated many essential features of the core genome (Lukjancenko et al, 2010), the extent and importance of genetic heterogeneity across populations (Avery, 2006), the ubiquity of horizontal gene transfer (Smillie et al, 2011), and the evolution and selection of functional genetic elements (David and Alm, 2011). At the same time, new computational tools have used the flood of data to model metabolic processes and signaling networks across the entire cell, generating many new testable hypotheses (Lewis et al, 2012). Most importantly, emerging advances in de novo synthesis and in vivo gene targeting allow empirical validation of these model‐driven hypotheses. By building and testing synthetic variants of biological systems, we have a unique opportunity to decipher the constraints imposed by the complexity of evolved systems and develop strategies for engineering living systems more conducive to quantitative modeling and rational design.

Here we review recent technologies that empower design‐based genome engineering approaches, identify potential bottlenecks, discuss strengths and limitations of strategies employing rational design versus evolution, and consider future applications of genome‐scale engineering. We advocate a synergistic engineering strategy that adopts the best aspects of rational genome design and evolutionary optimization.

Genome engineering is the art of constructing a genotype that gives rise to a desired phenotype, a challenge whose difficulty is influenced by the scale of genomic alteration required. One measure of scale is the number of changes that must be made to an existing genome to produce the desired phenotype. In some cases, this may require editing only one gene, a task that is clearly not genome scale. The same is true for a library of single‐gene variants and even a complete collection of single‐gene knockouts (Giaever et al, 2002; Baba et al, 2006), as each genome has only a single change. We define genome‐scale engineering to be any endeavor involving sequence modifications to at least two distinct regions of a genome. In what follows, we will mainly focus on technologies potentially capable of modifying large fractions of a single genome.

Genome‐scale engineering allows us to experimentally probe deep biological questions such as essentiality (Koonin, 2000), epistasis (Chou et al, 2011; Khan et al, 2011), encoding (Itzkovitz and Alon, 2007), evolvability (Tokuriki and Tawfik, 2009; Wagner and Zhang, 2011; Hill and Zhang, 2012), and robustness (Bershtein et al, 2006). At the same time, we aim to rationally build useful organisms that cannot be easily generated by harnessing evolution alone. Such endeavors require foundational tools in design, modeling, construction, and testing that extend from individual cells to populations of organisms (Figure 2). Iterations of design, model, build, and test phases are likely to be more important as the scale of the endeavor increases because biological complexity can grow exponentially. Below, we describe key features of these phases in genome‐scale engineering, outline current capabilities, and suggest opportunities for improvement.

Design is a set of specifications intended to achieve a dedicated objective under various constraints. Biological designs are those that describe the underlying blueprint of living organisms, built upon the information encoded in genes across the genome. As the focus of biological engineering shifts from individual genes to entire genomes, there is a growing need for more sophisticated genome design tools to assist such large‐scale engineering endeavors. Recordkeeping software is essential for tracking numerous modifications designed and generated across libraries of genomes. Traditional gene editors such as Vector NTI and SeqBuilder are largely inadequate for such purposes. However, new design tools and software suites such as J5 (Hillson et al, 2012), Clotho (Xia et al, 2011), and Genome Compiler ( http://www.genomecompiler.com/) provide better data management and user interfaces for the design of large operons and whole genomes.

Although recordkeeping is important, it is only one aspect of design, which must carefully define the experimental objective and triage candidate implementations according to likely failure modes. The complexity of biological systems often renders effective design a challenge. Fortunately, computational models can provide a useful guiding framework. Constraint‐based reconstruction and analysis (COBRA) models such as flux‐balance analysis have served as excellent predictive tools improve designs. These models generally rely on steady‐state analysis of metabolic flux to determine useful genomic targets that optimize a desired phenotype in silico. Although a detailed discussion of such models is beyond the scope of this review, COBRA‐based approaches have been reviewed extensively elsewhere (Lewis et al, 2012).

Whereas specialized metabolic models have been used for some years, Karr et al (2012) recently published the first complete virtual model of a cell, M. genitalium. At only ∼525 genes, M. genitalium is one of the smallest genomes known. Nevertheless, its phenotype is determined by the interaction of so many molecular components that it cannot be accurately modeled using any single method. To surmount this problem, Karr et al (2012) partitioned Mycoplasma into 28 distinct modules, modeled each using the most appropriate representation, and integrated the results to describe the entire cell. Analysis of unexpected behaviors on the part of the resulting virtual cell led to novel hypotheses concerning emergent controls on cellular behavior and identification of promiscuous enzyme activities capable of compensating for the lost genes. Despite these successes, accurate genotype‐to‐phenotype predictions of multiple genomic perturbations are still challenging due to biological complexity, large combinatorial variations, and computational limitations. Nonetheless, these examples demonstrate the power and utility of predictive models in understanding cellular behavior and identifying promising biological designs.

A complementary alternative to in silico prediction is direct experimental perturbation to identify potential targets and failure modes. Recent breakthroughs combining large‐scale mutagenesis with DNA sequencing have contributed significantly to improved genomic designs. Hutchison et al (1999) showed that sequencing transposon‐generated libraries of mutants can be used to systematically identify essential genes within the Mycoplasma genome. More recent approaches have employed next‐generation sequencing, including Insertion Sequencing (IN‐Seq) (Goodman et al, 2011), transposon sequencing (Tn‐seq) (van Opijnen et al, 2009), high‐throughput insertion tracking by deep sequencing (Wong et al, 2011), and transposon‐directed insertion‐site sequencing (Eckert et al, 2011). IN‐Seq, for example, involves the generation of libraries by random insertion of a Himar1 transposon containing a modified inverted repeat (IR) sequence. This IR is also recognized by the Type IIS restriction enzyme MmeI, which cuts the DNA 17 bases outside of its recognition site. When digested in vitro, genomic DNA carrying transposons harboring MmeI sites will generate fragments that include an extra 16–17 bp of genomic DNA, allowing high‐throughput sequencing to pinpoint the locations of all insertions. By enabling researchers to compare the abundance of individual mutants in the library before and after an experimental perturbation, Tn‐seq techniques enable multiplexed functional analysis of entire genomes. Every gene essential for the survival of a species can be identified in a single experiment that simultaneously rank‐orders all nonessential ‘accessory’ genes by their relative importance to organismal fitness under the conditions of interest. Other approaches such as global transcription machinery engineering (Alper and Stephanopoulos, 2007) and genome‐scale profiling of barcoded mutant libraries (Warner et al, 2010) can have a similar role in informing design. Expansion and broader adoption of these methods to guide genome‐scale design is needed for both single‐cell and multicellular organisms.

A wide variety of tools for targeted gene disruption and transgenesis are currently available (Figure 3). These tools vary considerably in their targeting efficiency, ease of retargeting, and effectiveness across a variety of different organisms (Table I). We focus on those with the greatest potential to enable large‐scale changes to single or multiple genomes by replacing large contiguous sequences or modifying numerous smaller sites serially or in parallel.

Empirical testing and validation of modified and synthesized genomes is necessary to determine whether the design goals have been met. Applying high‐throughput sequencing to confirm that a constructed genome matches its intended sequence is one such crucial test. Although viability and growth are also essential phenotypic tests, most design objectives require validation of function through other indirect assays. Moreover, many genome construction approaches result in libraries of different variants that require systematic curation to identify and isolate the best genomes from the rest of the population. Typical assays can be divided into low‐throughput and high‐throughput screens, which identify variants from populations of limited size (up to ∼10⁵), and high‐throughput selections, which enable the isolation of variants from much larger populations (Figure 5). For example, validating a constructed genome sequence by high‐throughput sequencing is a form of low‐throughput screen, while a viability assay testing the ability to survive and replicate under specific conditions is a selection. In both cases, the stringency of the assay is crucial, as constructs that do not generate the desired phenotype but still pass the screen or selection can lead to substantial delays and wasted effort. Selections are considerably more powerful when it is possible to generate large libraries of variants, as testing more variants increases the likelihood of finding ones with the desired phenotype.

Unfortunately, many desirable phenotypes cannot be directly selected, including small‐molecule biosynthesis and other traits that are among the most frequent targets for biological engineering. Low‐throughput screens can generally perform much more detailed phenotypic measurements by employing microscopy, transcriptomics, proteomics, or metabolomics to interrogate biological function at the cellular level. As our ability to build large libraries of genome variants grows, methods to increase the scale and throughput of such phenotypic measurements toward high‐throughput selections will be urgently needed to isolate and validate engineered genomes.

The application of genome‐scale approaches to metabolic engineering provides an excellent example of an integrated platform that showcases the synthesis of rational design, computational modeling, and multiplexed construction and testing to tackle real‐world biological engineering challenges. Numerous studies have used metabolic engineering to modify microbes to produce industrially relevant biochemicals and biofuels such as ethanol (Ingram et al, 1998) and higher alcohols (Atsumi et al, 2008), fatty acids (Steen et al, 2010), amino acids (Leuchtenberger et al, 2005), shikimate precursors (Bongaerts et al, 2001), terpenoids (Martin et al, 2003), polyketides (McDaniel et al, 1999; Pfeifer et al, 2001), and polymer precursors (e.g., 1,4‐butanediol (Yim et al, 2011)). A great example of genome‐scale metabolic engineering is Dupont's near‐decade long optimization of E. coli for bioproduction of 1,3‐propanediol (Nakamura and Whited, 2003). The industrially optimized strain required up to 26 genomic changes including insertions, deletions, and regulatory modifications. Recent advances in constraint‐based modeling (Lewis et al, 2012) have enabled in silico prediction of genomic targets whose perturbation may enhance strain performance or product yield. These computational predictions are ripe for experimental validation using new genome engineering tools. For example, OptKnock (Burgard et al, 2003), a computational tool that uses bi‐level metabolic flux optimization to predict the phenotype of gene knockout combinations, has been used to improve microbial production of lactic acid (Fong et al, 2005). Deleting different combinations of four identified genes (adhE, pta, pfk, glk) in E. coli significantly improved secretion of the desired product. Similarly, Alper et al (2005) described a set of strains generated through model‐driven combinatorial gene deletions of seven genomic targets that exhibited improved lycopene production by up to 8.5‐fold. More recently, Xu et al (2011) described the use of genome‐scale metabolic network modeling to generate genetic modifications that enhanced production of the useful precursor malonyl‐CoA. Knockout and overexpression genotypes in up to nine genes were generated combinatorially, with some strains containing up to five modifications (triple knockout, double overexpression).

Although these few studies suggest the promising potential of higher‐order mutants to access phenotypes needed to meet challenging design goals, the experimental difficulty of constructing such mutants has limited their use. The recent development of multiplex genome‐scale engineering tools such as MAGE has dramatically reduced the time required to generate combinatorial libraries of targeted mutations. We have shown that combinatorial exploration of both translation efficiency and gene deletions in up to 24 genes can yield useful combinations of genomic modifications for production of lycopene (Wang et al, 2009). More recently, the MAGE approach was extended to build a combinatorial library of genomic variants that contained synthetic T7 promoters in up to 12 genes involved in aromatic amino‐acid biosynthesis (Wang et al, 2012b). The combination of improved metabolic models and new techniques enabling combinatorial exploration and selection of specific genetic perturbations will substantially accelerate metabolic engineering (Sandoval et al, 2012).

When it comes to ease of designing, constructing, and testing genomes, not all organisms are created equal. Some have smaller genomes and unicellular lifestyles, while others have larger genomes and undergo complex multicellular development, both of which render genome design and modeling difficult. Some have many more tools available for genome editing, while others are burdened with polyploid genomes that increase the difficulty of constructing and testing new designs. Some organismal phenotypes can be readily measured, while others are subtle and hard to quantify. Most importantly, some replicate in mere minutes and are readily grown in large numbers, while others require years of labor‐intensive care to reach adulthood. The advent of new technologies for genome design, construction, and testing have compensated for some of these differences, but accentuated the impact of others.

Dairy cows are classic examples of slow‐growing, expensive, multicellular organisms that nonetheless have a large industry invested in their improvement. While cows have been modified through evolutionary engineering since antiquity, their slow growth and large diploid genomes render them recalcitrant to targeted variant construction and testing. Furthermore, in silico predictive models of mammals do not exist. Nevertheless, milk production has quadrupled over the last 60 years because the industry rigorously measured outputs and applied extremely strong selection in the form of artificial insemination (Funk, 2006). For decades, top bulls have routinely sired tens of thousands of offspring, efficiently transmitting only the best genes to the next generation—a purely blind evolutionary search, but the most effective strategy available given the constraints of the organism at the time. Thanks to high‐throughput sequencing, it is now possible to design strategies to accelerate the rate of improvement. Although we are far from understanding the mechanistic basis of milk production, recent genotyping sequencing efforts have begun to identify the chromosomal regions and individual genes favored by the past few decades of selection (Larkin et al, 2012). The industry is now implementing rationally designed generations‐long strategies to hasten the combination of known beneficial alleles into single genomes using selective breeding and perhaps, eventually, targeted genome editing.

Microbes are the mirror image of domesticated animals in almost every way. Unknown in antiquity due to their microscopic size, they tend toward small haploid genomes that can be grown quickly and in large numbers. Combined with a powerful selection, these traits permit swift evolutionary engineering, as first demonstrated by W.H. Dallinger's nineteenth‐century‐directed evolution of microbial thermal tolerance from 18°C to an astonishing 70°C over 7 years (Dallinger, 1887). A dearth of screening and selection technologies impeded further microbial engineering until the latter half of the twentieth century, but the subsequent explosion of such methods has rendered microbes—which combines rapid growth, large population sizes, and powerful selections—the organisms of choice for directed evolution studies. We recently demonstrated that even smaller and faster‐replicating genomes can further accelerate and even automate evolutionary engineering (Esvelt et al, 2011). Our system harnesses filamentous phages, which require only minutes to replicate in host E. coli cells, to optimize phage‐carried exogenous genes in a handful of days without researcher intervention. Compounding their growth advantage is the fact that microbes and phages are also ideal subjects for biological design, modeling, targeted genome editing, and genome synthesis, all of which can focus subsequent evolutionary searches on the regions of sequence space most likely to encode desirable phenotypes. Alternatively, these methods can compensate for the lack of a powerful selection that precludes evolution. Future technologies will ideally extend some of the advantages enjoyed by model organisms, such as E. coli and S. cerevisiae to other organisms, enabling more genome engineering endeavors to combine model‐driven targeted manipulation with the best growth and selection paradigm available to the target organism.

One of the overarching goals of genome‐scale engineering is to develop insights and rules that govern biological design. Unfortunately, most biological systems are riddled with remnants of historically contingent evolutionary events—a complex, highly heterogeneous state woefully unsuitable for precise and rational engineering. Rational genome design would be greatly facilitated by the construction of an underlying biological ‘chassis’ that is simple, predictable, and programmable. From that foundation, we can begin to build more complex systems that expand the repertoire of biochemical capabilities and controllable parameters. Furthermore, the chassis organism must contain mechanisms ensuring safe and controlled propagation, with strong barriers preventing unintended release into the environment and mechanisms that genetically isolate it from other organisms. The chassis should also contain obvious and permanent genetic signatures of its synthetic origins for surveillance of its use and misuse. Here we outline several classes of capabilities that should serve as a framework for a flexibly programmable biological chassis (Figure 6). A combination of current and future genome engineering technologies will be needed to construct such an engineered system.

Thus far, we have only considered methods for engineering individual genomes in the laboratory. Similar and related techniques might be adapted to modify most or all of the individual genomes that together constitute a single species: the pan‐genome. There are important safety and ecological considerations to assess before attempting any such project. Nevertheless, the environmental impact of human activity has already effected vast changes across the genomes of a large fraction of species all across Earth. It may be worth considering approaches that might correct such problems and accomplish desirable changes in a more benign manner. For example, we might spread a modification conferring drought resistance through the many local cultivars of a crop plant, with each cultivar retaining its local adaptations and genetic diversity. Such an approach would likely be superior in yield and lower in ecological impact to one in which all such variants are replaced with monocultures cloned from a single laboratory‐modified plant. Similarly, human disease vectors such as mosquitoes might be engineered to resist pathogen transmission, which would be considerably cheaper and more ecologically friendly than heavy insecticide use. Several genome engineering tools might be used to address these challenges. Targeting the wild‐type locus with nucleases would catalyze DSB repair using the transgenic cassette as a template, effectively converting all heterozygotes to homozygotes. Conceptually similar ‘gene‐drives’ have proven effective in the laboratory (Windbichler et al, 2011). Alternatively, a site‐specific recombinase targeted to the wild‐type locus could exchange the ends of homologous chromosomes, moving the desired modification to the formerly wild‐type chromosome and leaving behind a toxin rendering the donor chromosome sterile. Unlike other methods, this approach could be limited to a finite number of ‘jumps’ by placing a limited number of recombination sites and toxins on the initial donor, thereby improving our control over the spread of the engineered genetic element. Meanwhile, traits might be driven through microbiomes by combining horizontal gene transfer mechanisms with transposon‐ or retrotransposon‐mediated gene insertion. Further advances in these areas scaled to the ecosystem level (Mee and Wang, 2012) may extend our genome engineering capabilities across the pan‐genome, although we emphasize that ecological and safety considerations should be thoroughly assessed before such technologies are deployed.

Recent technological advances have overcome many of the limitations and bottlenecks that have constrained genome‐scale engineering. The exponential decrease in cost of DNA sequencing has dramatically accelerated forward genomics while enabling sequence confirmation of synthesized and edited genomes. New methods are bringing down the cost of DNA synthesis at an even faster rate. Emerging technologies for gene insertion, multiplex editing, and large fragment assembly have dramatically expanded our capabilities in certain model organisms, but further enhancements and extension to other organisms and across species will be needed to extend our engineering capabilities to the ecological level. Similarly, improved in silico modeling capabilities are urgently needed to guide rational genome design and synergize productively with evolutionary optimization. Finally, we suggest that the construction of a flexibly programmable biological chassis may serve as a foundation and standard for synthetic biology. These and other ambitious endeavors will continue to challenge our capabilities as genome engineers and our competence as biological designers.

The authors declare that they have no conflict of interest.

KME acknowledges funding from the Wyss Institute Technology Development Fellowship. HHW acknowledges funding from the Wyss Institute Technology Development Fellowship and the National Institutes of Health Director's Early Independence Award (Grant 1DP5OD009172‐01). We thank N. Lewis, S. Kosuri, and G. Church for insightful discussions and critical reading of this manuscript.