Genomes by design

Abstract

Next-generation DNA sequencing has revealed the complete genome sequences of numerous organisms, establishing a fundamental and growing understanding of genetic variation and phenotypic diversity. Engineering at the gene, network and whole-genome scale aims to introduce targeted genetic changes both to explore emergent phenotypes and to introduce new functionalities. Expansion of these approaches into massively parallel platforms establishes the ability to generate targeted genome modifications, elucidating causal links between genotype and phenotype, as well as the ability to design and reprogramme organisms. In this Review, we explore techniques and applications in genome engineering, outlining key advances and defining challenges.

Canadian author and Nobel Laureate Saul Bellow observed that “a writer is a reader moved to emulation” (REF. 1). Biological ‘reading’ and ‘writing’ are founded on pivotal discoveries regarding the structure, interpretation and manipulation of DNA, and have spurred a new biotechnological era in which the analysis and synthesis of whole genomes is now possible. Fruitful scientific investigations in the 1960s and 1970s led to the discovery of two ultimately interconnected techniques. Chain-termination sequencing developed by Frederick Sanger and Walter Gilbert provided an increasingly facile approach for reading the contents of genetic material². Around the same time, investigations into bacterial resistance to bacteriophage infection led to the discovery of restriction endonucleases and DNA ligases³. Soon after, these natural mechanisms were used in recombinant DNA technologies to assemble DNA fragments controllably and reproducibly in the first examples of ‘copy-and-paste’ functionality.

Early large-scale genome modification techniques (for example, transposon mutagenesis⁴) sought to affect phenotypic traits involving complex components but were limited by their untargeted nature, as much of the diversified population contained mutations that were unrelated to the cellular process of interest. Thus, studying numerous targeted mutations across genomes has been largely intractable. Genome engineering utilizes rationality and precision in genome design and mutagenesis to enable rapid and controlled exploration of a phenotype landscape (FIG. 1a). The goals of this process are varied and include elucidating combinatorial genetic variants that contribute to organism behaviour, producing expanded phenotype topologies via new construction components such as non-natural biochemical building blocks and increasing evolutionary robustness⁵ (FIG. 1b).

Figure 1. Genes, networks and genomes by design.

a | Both Darwinian evolution and rational genome engineering rely on the introduction of mutations to traverse a genotype space and explore the resultant changes in cellular phenotype. The high-dimensional nature of genotype landscapes presents a key challenge in resolving causal genotype–phenotype relationships and engineering organisms with prescribed behaviours. Genome engineering seeks to leverage existing knowledge, modern DNA manipulation techniques and evolution to search through complex genotype landscapes more efficiently. b | The genome can first be thought of as a physical collection of its constituent DNA elements. These constituent parts can be conceptualized as nodes in a network in which edges represent functional connections (for example, genes involved in the same cellular process). Networks can range in size from small collections of genes in biosynthetic clusters to the genome-scale network of all genes and cellular processes. The space of all possible genotypes can be visualized as a phenotypic landscape where each genotype is associated with a particular level of performance with respect to various phenotypic traits. Manipulation of the genome can alter the connections in cellular networks or add new nodes. These new networks can represent a new location in the phenotype landscape or alter the phenotype landscape itself. Changes at the gene and small-network level enable exploration of broadened phenotype landscapes, whereas large-scale genome modification can yield disruptive and novel fitness landscapes.

Laboratory studies of global evolution lack control over mutational processes, which limits the specific genotype landscapes explored over the course of a single experiment⁶. Furthermore, evidence suggests that evolution is driven by the cumulative effect of multiple, moderately beneficial mutations rather than individual events that become fixed across a population⁷. Coupling these observations elucidates unique challenges for genome engineering efforts: accessing multiple mutations within a single generation and enabling researchers to bypass adaptive valleys observed in non-engineered evolutionary processes.

Different conceptual approaches exist to traverse phenotype landscapes⁸. For example, de novo DNA synthesis and assembly techniques leverage extensive in silico design phases to anticipate and build a limited number of episomal constructs for characterization. Alternatively, as the effects of large-scale genomic changes on fitness are often unpredictable, many technologies seek to explore a large combinatorial space of genomic alterations. Rather than direct phenotyping, these larger chromosomal libraries necessitate sorting or selections in which beneficial elements of genotypic diversity are propagated via their phenotypes⁹. Uniquely, genome engineering approaches can be utilized to sample a targeted, combinatorial genetic landscape in order to elucidate causal links between genotype and phenotype.

The microbial biome is a rich, largely untapped resource¹⁰ that has a central role in ecological homeostasis¹¹, and promises new medicinal compounds, routes to natural bioremediation and platforms for metabolic engineering. Owing to their relatively small genome sizes, well-described and malleable mechanisms of DNA repair, and fast reproduction rate, microbial genomes are favourable targets for genetic engineering. Genome sequencing and concurrent computational advances have established a knowledge base that now necessitates nucleotide resolution for functional genetic testing and reprogramming. A mutable, controllable genome is required to fully interrogate and manipulate the phenotype landscape. This Review is chiefly concerned with the rationale, technologies and applications of genome design. When coupled with next-generation sequencing (NGS), the iterative genome engineering process — design, build, test, evolve and learn — has become a powerful pipeline for the exploration of emergent and novel phenotypes.

Gene-scale engineering

Although many cellular properties are multifactorial and emerge from systems-level interactions encoded across many genomic loci, gene-scale engineering targets single genes contributing to a desired phenotype. Conceptually, we divide gene engineering efforts into two functions: the identification of a gene that is capable of eliciting a desired phenotype; and the targeted modification of a host genome. Directed evolution of single genes is a mature platform where a large, diverse library of sequences produced in vitro by methods such as DNA shuffling¹² and PCR mutagenesis¹³ are subsequently screened or selected for function. An inherent limitation of directed evolution strategies is that the very large sequence libraries may contain gene variants that are capable of driving the desired phenotype at very low frequencies. One solution to this challenge is to reduce the complexity of the sequence space by first designing sequences to test in silico¹⁴ and then synthesizing them de novo.

These techniques create static, large libraries of DNA sequences that require researcher participation for further diversification and multiple rounds of selection. Solving the challenge of continuous evolution promises to enable autonomous diversity production. One such approach was identified in the field of biomolecule evolution in the form of phage-assisted continuous evolution¹⁵. By linking the expression of phage protein III (pIII) to the activity of their target gene, Liu and colleagues¹⁵ formulated a selection to engineer and modify bimolecular interactions of interest. Through suppressed proofreading¹⁶ and increased error-prone lesion bypass¹⁷, the authors suggest that the entirety of a random single- and double-mutant genotype landscape may be explored over a single evolutionary generation. Although the original platform was limited to selection by affinity (that is, positive selection), a subsequent study identified a dominant negative mutant of pIII that could be used to drive a negative selection, enabling evolution of T7 RNA polymerase enzymes with altered substrate specificities¹⁸.

One of the unifying factors of these gene design and selection strategies is that diversified libraries are typically cloned to plasmids or phage. In contrast to chromosomal targets, episomal constructs have downsides, including biosynthetic burden, poor persistence over time¹⁹, highly variable expression²⁰ and an associated decrease in organismal fitness^20,21. Genome engineering of the host chromosome (or chromosomes) circumvents pleiotropic plasmid effects while also ensuring the maintenance of physiological gene copy numbers.

Nuclease-based genome editing

Nuclease-based genome editing approaches enable gene disruption via non-homologous end-joining or targeted editing using homologous templates (FIG. 2a). Both zinc-finger nucleases (ZFNs) and transcription activator-like effector nucleases (TALENs) are composed of an engineered DNA recognition domain fused to the FokI endonuclease catalytic domain^22–24. Engineered ZFNs are assembled by linking individual modular domains with empirically determined target specificities²³. TALENs utilize a similarly modular approach, but each domain contains repeat-variable di-residues that determine specificity in a highly reproducible manner²³. The key difference between these targeting methods is that each zinc-finger module is thought to associate with a three-base sequence, whereas TALENs have single-base resolution. Both ZFNs and TALENs are areas of fruitful, ongoing investigation^24,25.

Figure 2. Nuclease-mediated genome engineering.

a | Nuclease-based gene editing using zinc-finger nucleases (ZFNs), transcription activator-like effector nucleases (TALENs) or CRISPR–Cas targeting induces double-stranded breaks that are repaired in a template-dependent (homology-directed repair) or -independent (non-homologous end-joining) manner. b | Modified CRISPR–Cas systems have been used as transcriptional activators by stabilizing RNA polymerase (RNAP) binding at specified sites. c | Transcriptional repressors that use modified CRISPR–Cas to block transcriptional initiation or elongation have also been developed. d | Targeting CRISPR–Cas cleavage to sequences that are unique to a specific genome results in the selective removal of that genome from a bacterial population. e | In the context of recombination-mediated genome engineering, targeting CRISPR–Cas cleavage to the wild-type sequence has been used to remove unaltered cells and enrich for mutations. f | Gene drives using CRISPR–Cas systems cut homologous chromosomes and promote their own incorporation via homology-directed repair. Placement of constructs near genes of interest can facilitate their spread throughout a population. PAM, protospacer adjacent motif.

The emergence of CRISPR–Cas²⁴ is rapidly changing the field of microbial genome editing (FIG. 2b,c). CRISPR–Cas9 is a bacterial adaptive immunity system that genomically stores foreign DNA in response to infection²⁶. In type II CRISPR–Cas, arrays of foreign DNA are transcribed and processed into CRISPR–Cas9 RNAs (crRNAs), annealed to trans-activating crRNA (tracr-RNA) and complexed with the Cas9 endonuclease. This complex can then target foreign genetic material for cleavage²². It was later shown that the crRNA and tracr-RNA can be engineered into a single construct (known as guide RNA (gRNA)) that is sufficient for targeted nuclease activity²⁷. The endonuclease complex cleaves a 23 bp target sequence comprising 20 bases of guide sequence and the required protospacer adjacent motif (PAM)^24,28. Unlike target specificities for ZFNs and TALENs, target specificity for CRISPR–Cas is derived from the gRNA sequence rather than protein modules, and CRISPR–Cas9 systems are easily modified for, in principle, any genomic target proximal to a PAM. Numerous improvements have already been described to address off-target cleavage events²⁹, including the use of paired nicking CRISPR–Cas9 (REF. 30) or FokI³¹. CRISPR–Cas9 has been shown to enable direct editing in many settings^22,24, including human cells^32–34. Furthermore, applications in transcriptional regulation^35,36 (FIG. 2b,c), targeted bactericidal activity^37,38 (FIG. 2d,e) and gene drives^39,40 (FIG. 2f) have emerged rapidly. Here, we focus on two distinct genome engineering challenges: combinatorial genome editing and small chromosomal library insertion.

The natural function of CRISPR–Cas9 in adaptive bacterial immunity suggested that multiple encoded guide sequences might be tiled into a single array for multiplexed editing of a single genome. In their pioneering study, Cong et al.³⁴ showed deletion of two targets in eukaryotic cells with a single array. Furthermore, a study in Saccharomyces cerevisiae reported 43% and 19% efficiencies for disrupting loss-of-function mutations when targeting two and three sites, respectively⁴¹. Although earlier examples of multiplexed editing used individual gRNA for each target site, this approach has several draw-backs, including transfection of multiple, rapidly diluted plasmids, and variable copy numbers⁴². One study used a single lentiviral vector expressing Cas and up to four gRNAs to convert four loci in three out of ten cells⁴². More recently, CRISPR–Cas9 complexes were used to transcriptionally activate up to ten sites in a controlled manner⁴³, simultaneously showing that pooling of gRNAs led to decreased efficiencies. Although this result has not been recapitulated in the context of genome editing, these findings suggest that CRISPRs may have increased multiplexing capacity for gene disruption. However, organism-dependent upper limits on double-stranded breaks and nicks are currently unknown.

In addition to direct editing of arbitrary genomes, CRISPR–Cas9 may enable the use of chromosomal, rather than episomal, gene libraries. As discussed, plasmid libraries for gene-scale engineering have several limitations, whereas homologous recombination is too inefficient for DNA library insertion and requires numerous antibiotic markers. In a recent study, Ryan et al.⁴¹ reported that the use of error-prone PCR to create a library of gene mutants, followed by CRISPR–Cas9-mediated insertion into the S. cerevisiae genome, improved consumption of the industrially relevant compound disaccharide cellobiose. This report highlights the potential synergies between CRISPR–Cas9 gene editing and genome engineering in organisms with well-described mechanisms of homologous recombination. Given the rapid developments in CRISPR–Cas9 technologies, we anticipate numerous imminent advances that will improve upon mutation efficiency, expand the already broad range of substrate organisms and increase throughput as required to further develop these platforms for large-scale genome engineering. Furthermore, proof-of-concept multiplex CRISPR–Cas9 studies have largely addressed gene disruption, suggesting that efficient, high-throughput homology-directed repair with provided templates requires further investigation.

Recombineering for gene and network design

Whereas nuclease-based engineering technologies leverage DNA-binding domains to confer target specificity, recombineering (recombination-mediated genetic engineering) relies on homologous recombination to recombine single-stranded DNA (ssDNA) or double-stranded DNA in the genome (FIG. 3a). In seminal work, it was shown that functional replacement of Escherichia coli RecBCD with the phage lambda Red operon (including the Beta, Exo and Gam genes) yielded a strain that was capable of recombining short, linear DNA fragments at a significantly improved rate^44,45. As a result, homologous recombination was made accessible by the use of primers and facile amplification with PCR. It has been proposed that the oligonucleotides act as synthetic Okazaki fragments and are incorporated in the lagging strand of the replication fork^44,45. Subsequent work by Costantino and Court⁴⁴ improved the recombination rate at a single locus by >100-fold by removing E. coli mismatch repair.

Figure 3. Gene and network-scale recombineering.

a|Recombineering (recombination-mediated genetic engineering) uses cellular homologous recombination enhanced by phage proteins to recombine single-stranded DNA (ssDNA) or double-stranded DNA into the genome. This enables site-specific introduction of mutations ranging in resolution from single base pair changes to the introduction of entire genes. b | The SCRIBE (Synthetic Cellular Recorders Integrating Biological Events) analogue memory system uses reverse transcriptase (purple) to produce ssDNA in vivo, which is then used in recombineering. SCRIBE is able to introduce heritable genetic changes (*) in response to a predetermined stimulus. These heritable changes act as a cellular memory of exposure to the stimulus. The fraction of cells containing this mutation serves as an approximation of the cumulative stimulus input function. c | Multiplex automated genome engineering (MAGE) uses the introduction and incorporation of synthetic ssDNA at the replication fork to introduce desired mutations. Mutations can be targeted to multiple residues in a single gene, multiple genes in a pathway, or across the genome. d | Trackable multiplex recombineering (TRMR) utilizes recombineering and high-throughput profiling to introduce mutations and to identify variants that provide a fitness benefit under different conditions. Each recombineering cassette contains a selectable marker, targeting homology domain, and a sequence to increase or decrease gene expression. Sequencing of genetic barcodes associated with each mutation reveals the resultant enrichment of mutations that confer a fitness benefit.

Recombineering with the lambda Red genes and homologues has been extended from E. coli to numerous organisms^46–48, and has been combined with CRISPR–Cas9 and gRNA to enable 65% mutation efficiency in E. coli ⁴⁹. By utilizing a dual-RNA–Cas9-directed cleavage to kill unmated cells, the authors showed that direct selection against non-edited cells is a powerful approach for efficient recovery of mutated genomes. Although this method has a high probability of yielding a desired genotype, it yields relatively few surviving clones, limiting potential library size. Notably, the same study observed nearly 100% CRISPR–Cas9 efficiencies in the highly recombinogenic Streptococcus pneumoniae, again suggesting⁴¹ that the host recombination machinery may prove to be a crucial adjuvant for CRISPR–Cas9 in microorganisms in which recombineering is not possible.

Recently, recombineering was utilized to create an analogue memory system in vivo⁵⁰. The SCRIBE (Synthetic Cellular Recorders Integrating Biological Events) approach uses an encoded reverse transcriptase to produce a hybrid RNA–ssDNA molecule from an arbitrary provided sequence (FIG. 3b). When utilized alongside the Beta protein, this approach drove recombination at approximately 10⁻⁴ events per generation. By linking the expression of these components to inputs such as small-molecule inducers or optogenetic control, the authors describe a cumulative effect function in which the proportion of the population containing a mutation is proportional to the total cellular input signal, regardless of distribution over time. We anticipate that the key innovation presented in the SCRIBE study — that is, the use of reverse transcriptase to drive in vivo oligonucleotide production coupled with lambda Red recombination — will provide myriad downstream applications in genome engineering. For example, error-prone RNA polymerases⁵¹ or reverse transcriptases⁵² may enable the production of diversified oligonucleotides for autonomous in vivo evolution. This study also suggests that multi-site continuous recombination is possible, but will require concomitant increases in efficiency.

Network design technologies

Chromosomal network engineering

Recombineering technologies laid the groundwork for parameter optimization and the ability to target multiple genomic loci for oligonucleotide-mediated allelic replacement (FIG. 3c). Inspired by advances in ssDNA recombineering, multiplex automated genome engineering (MAGE) uses lambda Red recombination and synthetic phosphorothioate-protected oligonucleotides to introduce targeted mutations across the genome^53,54, enabling user-defined indels or mismatches with efficiencies of ≤30%. Although MAGE was developed with the goal of whole-genome redesign and recoding of E. coli (as discussed below), stochastic incorporation of oligonucleotides across multiple genetic loci enabled the rapid creation of combinatorial genetic diversity within a cell population and facilitated its immediate use for pathway diversification. In a proof-of-concept demonstration of targeted, multi-site genome diversification, MAGE was used to optimize the 1-deoxy-d-xylulose-5-phosphate (DXP) pathway for lycopene production. The ribosome-binding sites of 20 enzymes in the DXP pathway were targeted using oligonucleotides containing partial degeneracy biased towards Shine–Dalgarno consensus sequences, enabling the identification of optimal cellular production from the diversified population. Additional nonsense mutations were targeted to four enzymes in competing pathways in order to maximize flux through the DXP pathway. The combination of these efforts yielded strains that were capable of up to a fivefold increase in lycopene production and, through comparison of mutations identified by sequencing, the authors elucidated the causal genetic basis for improved phenotypes. Subsequent work has shown that selections can dramatically increase observed allelic-replacement frequencies^55,56 and have been used in the context of promoter engineering⁵⁷. DNA microarrays can be a facile oligonucleotide source for lambda Red genome engineering⁵⁸, as shown in a proof-of-concept study that created a library of ~13,000 130 nt ssDNA oligonucleotides (BOX 1). These experiments highlight the importance of creating combinatorial genetic diversity and subsequent characterization when optimizing network phenotypes.

Box 1. Large-scale DNA synthesis and assembly.

Advances in Sanger sequencing and next-generation sequencing (NGS) have enabled reading of genomic texts, whereas de novo DNA synthesis serves as the author’s ink. Solid-phase phosphoramidite chemistries that developed in the 1980s are the workhorses of column-based oligonucleotide synthesis methods, enabling the production of short single-stranded DNA (ssDNA) with moderate parallelization (≤384 oligonucleotides)¹⁹. New, large-scale DNA microchip-based synthesis methods look to provide a scale that parallels NGS more closely^129–131 (FIG. 6).

DNA sourced from microarrays permits high-density synthesis of ~10⁵ customized oligonucleotides^129,132. Microarray synthesis yields approximately picomolar quantities, requiring PCR amplification of select pools of oligonucleotides. Two complementary approaches, barcoding with pool amplification¹³³ and microwell assembly¹³⁴ coupled with error correction^129,131, have enabled the use of microarray-based oligonucleotide pools^129,135 as a low-cost source for large numbers of oligonucleotides. Gene synthesis, the overlap assembly of short oligonucleotides (~50–200 nt) into larger fragments, relies on an array of approaches, including PCR¹³⁶, circular amplification¹³⁷ and enzymatic assembly^138,139 (see the figure).

As shown in the figure, DNA microchip oligonucleotide synthesis can be coupled with either subpool PCR or microwell printing for the construction of intermediate-length constructs, which can then be assembled in vitro or in vivo for use in gene, network and genome engineering.

Building on a seminal study on microarray-based synthesis of translational genes¹³¹ and subsequent improvements in oligonucleotide quality, one study by Goodman et al.⁸¹ used a library of >14,000 microarray-synthesized combinations of promoters, ribosome-binding sites and amino-terminal codons to elucidate roles of codon bias in bacterial protein expression, showing that large synthetic DNA libraries can provide a platform for testing hypotheses about fundamental biological questions. De novo synthesis has become a viable and increasingly facile tool for the production of libraries of rational designer genes.

As efforts turn to bridging these advanced DNA synthesis capabilities with de novo synthesis of whole genomes⁷⁹, the ‘design, build and test’ engineering design cycle will require significant optimization. Synthesis affords complete control over new genome architectures, but informing the design process remains a challenge. Over the past 5 years, microarray-based synthesis and associated assembly technologies have increased quality and reduced the cost of large-scale gene synthesis, establishing preliminary progress on future multiplexed synthesis and assembly of refactored pathways⁶⁹. Efforts to scale towards synthesis of multiple genomes will require improved assembly technologies performed in vitro and in vivo with the aid of automation. Striking advances in rapid and facile phenotypic characterization will be essential for the testing of designer genomes.

The selection of MAGE targets requires previous knowledge of the biological processes under investigation. Trackable multiplex recombineering (TRMR) utilizes barcoded synthetic DNA cassettes to introduce expression-modulating mutations across the genome, enabling the identification of fitness-improving genes under previously uncharacterized biological or environmental conditions⁵⁹ (FIG. 3d). Each cassette is designed to contain a selectable marker, homology domains that target it to an individual gene in E. coli, and a sequence designed either to increase or to decrease gene expression. With this system, Gill and colleagues⁶⁰ were able to generate a mutant population in which more than 95% of E. coli genes had modified expression. The authors illustrated the utility of TRMR by exposing the mutant population to various environmental conditions and measuring the differential abundance of mutants before and after the population exposure. The abundance of each mutant was determined by sequencing the barcodes present in the two populations, subsequently enabling identification of genes that improved fitness. In one experiment, TRMR revealed several mutations that enabled cells to better tolerate lignocellulosic hydrolysate, including mutations affecting vitamin metabolic processes and in the gene encoding adenylate cyclase⁶⁰. This broad, genome-scale approach can help to establish novel connections between genes and cellular fitness.

Further advances in multiplexed recombineering are still required, as the platform has only been explored in a select number of hosts with variable efficiency^61,62, and ongoing researcher input is required to continue cycling cells through the experimental pipeline. Recent work has enabled oligonucleotide-based genome engineering alongside active mismatch repair machinery through the use of modified nucleotides, among other approaches⁶³. Future combinations of these multiplexed recombineering approaches and continuous in vivo oligonucleotide production via reverse transcriptases promise to dramatically increase the efficiencies and range of genomic alterations. Bacterial artificial chromosomes are known to tolerate large synthetic constructs⁶⁴ and may be capable of driving the production of >10³ oligonucleotides for continuous and autonomous multi-site genome modifications in vivo. Oligonucleotide-mediated engineering is a facile, inexpensive, and powerful approach for large-scale, in vivo genome manipulation.

Episomal network engineering

In many cases, it is desirable to import functionality from unculturable, slow-growing, or otherwise challenging organisms into a more facile host such as E. coli. One such example, the 32 kb 6-deoxyerythronolide B synthase gene cluster, provided an early proof of concept for the de novo synthesis of whole biological networks⁶⁵ (FIG. 4a). Genes in polyketide synthase (PKS) clusters are kilobases in size, and alteration of individual modules can lead to the biosynthesis of novel polyketides. To facilitate this exchange, a subsequent study introduced design elements to the PKS cluster in the form of unique restriction sites, enabling the assembly of 154 combinations of PKS clusters⁶⁶. These studies reflect synergies between bioprospecting (the systematic search for medically or industrially relevant pathways and small molecules), massively parallel genome sequencing and de novo DNA synthesis, enabling the importing and optimization of frequently minimally expressed pathways in genetically tractable hosts.

Figure 4. Refactored pathways and minimal genomes.

a | There are numerous approaches to the transfer of gene clusters. Biosynthetic clusters can be cloned from their native organism and transferred to another organism for heterologous expression. For organisms that are recalcitrant to laboratory culture techniques, gene sequences of interest can be identified from genomic or metagenomic databases and subsequently synthesized and introduced into a production organism. The native regulatory elements of a heterologous pathway may not be optimal for expression in the host organism. Refactoring approaches seek to rebuild the pathways using well-characterized modular regulatory elements and removing all native regulation. b | Recombineering (recombination-mediated genetic engineering) was used to delete up to 15% of the wild-type Escherichia coli K12 genome, including insertion sequence (IS) elements. This led to a more stable and streamlined genetic architecture that manifested itself in altered cellular phenotypes. c | The de novo synthesis of the Mycoplasma mycoides genome utilized synthetic DNA oligonucleotides assembled first in vitro and then in Saccharomyces cerevisiae. Following completion of the ~1 Mb genome, the whole construct was transplanted to Mycoplasma capricolum, yielding a viable organism. BAC, bacterial artificial chromosome; RBS, ribosome-binding site.

One appealing method to disentangle and control the components of a complex genetic circuit is to begin by refactoring the interwoven molecular parts into modular, defined units^67,68. This engineering goal is intended to facilitate inter-species transfer and, simultaneously, to enable exogenous control over the system⁶⁹. In a pioneering study by Endy and colleagues⁶⁸, nearly 30% of the T7 bacteriophage genome was replaced with engineered DNA, resulting in viable phage that was simultaneously simpler to model and manipulate. In recent work, Voigt and colleagues⁶⁷ used large-scale synthesis to address a long-standing challenge in understanding and optimizing the nitrogen fixation (nif) gene cluster in Klebsiella oxytoca for use in a facile host^67,69. Using a refactored nif cluster, the authors designed a combinatoric library of 3 promoters, 12 ribosome binding sites, 7 spacers and 6 genes that underwent codon optimization across a total of 84 constructs. They then introduced the constructs into a K. oxytoca strain deficient for the six genes and screened for function, recovering a number of well-performing clusters. They repeated this process iteratively across different subsets of the nif cluster as well as on the entire 16-gene pathway. This study was able to close the design loop by performing functional assays, RNA sequencing, collecting growth information and feeding these data back into the cluster design. Although transfer to E. coli led to modest functionality, the computational (‘design’), synthesis (‘build’), and experimental (‘test’) pipelines in this study suggest a broad range of applications for heterologous pathway redesign, characterization and optimization. This work will benefit from future applications of recombineering and network mutagenesis technologies^53,54,70 that introduce further genetic diversity to the synthesized pathways, effectively using de novo synthesis as a scaffold for in vivo evolution.

Although many of the discussed methods, from directed evolution to de novo pathway synthesis and refactoring primarily focus on the addition of gene function, recent developments in transcriptome engineering may provide the inverse functionality. Even though RNA interference (RNAi) is not utilized in bacteria, synthetic RNAs (sRNAs) composed of scaffold and target-binding sequences have been shown to be capable of strongly repressing gene expression in E. coli^71–73. Similar efforts in S. cerevisiae have used engineered RNA regulators^74,75 as well as RNAi^76,77 to enable rapid strain prototyping in yeast. Taken together, these transcriptome modification technologies have shown great promise for facile, rapid and scalable engineering in microorganisms.

Design of reduced and chimeric genomes

The commoditization of DNA sequencing and synthesis and the emergence of large-scale assembly (BOX 1) and mutagenesis techniques have enabled genome-scale design and manipulation. One long-term area of investigation rewarded by these advances is the search for a minimal microbial genome^64,78,79, which can then be used as an efficient foundation for next-generation applications. The first landmark in this effort was the transposon mutagenesis of Mycoplasma genitalium, which found that 55–73% of the 517 genes were essential⁸⁰, a figure later revised to 79%⁸¹. In related work, Blattner and colleagues⁸² utilized recombineering in E. coli K12 to investigate phenotypic outcomes of reduced genome size^5,6,82 (FIG. 4b). Rather than pursuing random mutagenesis, the authors used bioinformatics approaches to identify genome segments that were unique to K12 and then used recombineering to delete up to 15% of the genome, yielding a strain with increased transformation efficiencies. By specifically targeting insertion sequence (IS) elements and other mobile DNA elements that are frequently involved in recombination, the engineered strains displayed both increased genomic stability and the ability to maintain plasmids with deleterious genes.

The first whole-genome transplantation⁸³, complete chemical synthesis of the M. genitalium chromosome⁶⁴, and the synthesis followed by transplantation of the Mycoplasma mycoides genome⁷⁹ marked important proof of concepts in genome engineering^64,79 (FIG. 4c). In the case of the ~1 Mb M. mycoides genome, 11 ~100 kb assemblies were built from short DNA oligonucleotides using overlap assembly and then homologous recombination in S. cerevisiae. In parallel, constructs containing 1 of the 11 assemblies and supplemented with the remainder of the wild-type M. mycoides genome were used to evaluate the function of individual components. Although the first-stage intermediates were sequence-verified, one ~100 kb assembly contained a single point mutation in dnaA that precluded transplantation. Upon repair of this mutation and reassembly of the genome, the team was able to successfully transplant and sequence the M. mycoides genome in Mycoplasma capricolum.

This remarkable series of studies suggests that it may be possible to produce functional cells from pre-determined sequences. At the same time, because the de novo synthesis strategy relies on assembly outside the target cellular context, design errors or point mutations may dramatically alter the desired organismal phenotype. The cloning of a chimeric Synechocystis PCC6803 and Bacillus subtilis genome¹⁴⁰ suggests that attention to design constraints (for example, a single ribosomal RNA species for each genome) or key culturing parameters may be required for genome construction efforts. Indeed, the synthesis of larger and more complex genomes may reveal otherwise masked epistatic relationships. The complete syntheses of Mycoplasma genomes and others⁸⁴ have inspired genome engineering based on DNA synthesis and, when combined with other design principles such as refactoring, may deliver ‘plug-and-play’ production-optimized genomes.

Genomic recoding and orthogonal biology

Genetic code expansion

A central goal of genome-scale engineering is the ability to fundamentally redesign genomes to explore and uncover novel phenotypic landscapes. Questions regarding the redundancy, mutability, and structure of the genetic code are challenging to address⁶, and reconstruction of alternative genetic codes was, until recently, infeasible. Historically, genetic code expansion has the aim of increasing the biochemical diversity of translated proteins. Beyond the 20 standard amino acids, some species are able to utilize selenocysteine⁸⁵ and pyrrolysine⁸⁶, whereas further chemical space is accessible to many organisms via post-translational modifications (PTMs). Synthetic chemistry has enabled in vitro access to a wide variety of compounds, but these are not necessarily accessible to the replicating cell. Thus, with the vision of creating biological foundries and organisms with novel properties, considerable efforts have been extended towards the expansion of the cellular amino acid repertoire⁸⁷.

Early studies recognized that chemically aminoacylated tRNA⁸⁸ or amino acid auxotrophs⁸⁹ could be used to drive utilization of a non-standard amino acid (nsAA). It was quickly recognized that in vivo genetically encoded translation using a nsAA would require the design of orthogonal translation systems (OTSs)^90,91. An OTS is composed of a tRNA that does not interact with native aminoacyl-tRNA synthetases (aaRSs) and an aaRS that specifically and uniquely charges the tRNA with a nsAA. Furthermore, the orthogonal tRNA must then suppress a codon that does not otherwise encode a standard amino acid. The first such example in vivo utilized the tyrosyl-tRNA aaRS and tRNA from Methanococcus jannaschii to encode O-methyl-l-tyrosine in response to the amber stop codon TAG⁹². Although other methods including quadruplet-decoding ribosomes have found successful application^93,94, amber suppression has become the most common method of genetic code expansion, partially because it is the least abundant codon across the genome.

Amber suppression has enabled the DNA template-controlled incorporation of diverse chemistries by >100 nsAAs, including tyrosyl-derivatives⁹⁵ (see REF. 91 for an extensive list), and PTMs^96,97. The multiple instances of TAG codons throughout the genome lead to uncontrolled proteome-wide nsAA incorporation, affecting cellular fitness. Furthermore, competition between the amber-suppressing tRNA and release factor 1 (RF1) limits the production of nsAA-containing polypeptides. For this reason, numerous efforts were made to alter both RF1 and RF2 (REFS 98,99) or to delete RF1 and supply essential TAG-terminating genes episomally¹⁰⁰. These findings were extended by the discovery that the conversion of seven TAG codons to TAA in essential genes enabled deletion of RF1 and a subsequent increase in protein yields at the cost of cellular fitness⁹⁷. One consequence of this process was the rapid evolution of glutamine amber suppressors, possibly relieving ribosomal stalling at more than 300 remaining TAG sites¹⁰¹.

Genomic recoding

The confluence of numerous technologies has enabled the first genome-wide replacement of a single codon and reassignment as an open coding channel¹⁰¹ (FIG. 5a). Recognizing that the genome of E. coli contains 321 TAG codons, MAGE^53,54 was used to convert TAG codons to TAA in 32 separate strains¹⁰². By maintaining evolutionary context in vivo, each set of mutations could be evaluated for viability and phenotypic effects. As the E. coli genome was being edited, rather than transplanted⁷⁹ or incorporated^103,104, a new method for conjugative assembly genome engineering (CAGE)^102,105 was developed. CAGE utilizes conjugation to hierarchically combine distinct genotypes from engineered E. coli into a single strain, enabling compilation of four partially recoded strains. In a subsequent study, it was shown that full genomic recoding enabled RF1 deletion and reassignment of the TAG codon to incorporate nsAA¹⁰¹. This process was without observable pleiotropic effects associated with genomic recoding (for example, the emergence of tRNA suppressors).

Figure 5. Genome recoding and new biological functions.

a | The genomically recoded organism was constructed by first using multiplex automated genome engineering (MAGE) to convert TAG stop codons to TAA in a panel of strains. Conjugative assembly genome engineering (CAGE) was then used for the hierarchical assembly of these partially recoded genomes into a fully recoded genome in which all 321 TAG stop codons are reassigned. Release factor 1 (RF1) was then deleted in order to establish TAG as an open codon for the incorporation of non-standard amino acids (nsAAs). b | The addition of an orthogonal translation system capable of charging a nsAA to a TAG-recognizing tRNA enables the genetic encoding of nsAA-incorporating proteins. c | The reassignment of TAG from a stop codon can render TAG-containing or -terminating genes non-functional in a different organism. This reduces the probability of horizontal gene transfer (HGT) between recoded and natural organisms. d | The incorporation of TAG codons in essential genes provides a method for the biocontainment of recoded organisms by linking its viability to the presence of an exogenous nsAA. aaRS, aminoacyl-tRNA synthetase.

The resultant genomically recoded organism (GRO) reveals a dramatically altered phenotypic landscape. As there are no TAG codons present in the genome, nsAA incorporation via amber suppression is exclusively directed by the DNA templates provided and uncontested by RF1, leading to high purity and yields of diverse proteins (FIG. 5b). Second, the recoded E. coli exhibits increased resistance to T7 bacteriophage, revealing that genome recoding may interfere with the function of genes transferred from viruses, between species and across ecosystems (FIG. 5c). Genes ending with TAG that are transferred into the recoded E. coli will not be translated properly, and TAG-containing genes transferred from this strain will result in premature truncation. These proof-of-concept horizontal gene transfer experiments suggest that further recoding^101,106 will enable more stringent genetic isolation and multi-virus resistance.

The long-term evolutionary outcomes of genetic code expansion in E. coli or other hosts are as yet unknown⁶. Despite deep conservation of the genetic code, there is evidence of TAG and TAG codon reassignments in bacteriophages and TAG reassignments in bacteria¹⁰⁷. A recent study involving exposure of phage to 3-iodotyrosine-incorporating E. coli revealed that the phage mutated to preferentially incorporate the nsAA in the lysis-involved protein type II holin, yielding an improved evolutionary phenotype¹⁰⁸. Similar experiments in E. coli exploring evolution in the context of non-natural biochemical building blocks^109,110 may require the development of OTSs with improved functionality as current platforms are several orders of magnitude less efficient than native aaRS–tRNA pairs^95,111–113, providing a potential counter-selection to widespread nsAA incorporation.

Recoding of other codons appears to be a plausible aim, as evidenced by a study that utilized DNA microchips to investigate removal of 13 rare codons from a panel of 42 highly expressed essential genes¹⁰⁶. This report observed that synonymous codons used at a single locus within an essential gene yielded unpredictable variability, reinforcing the importance of a rapid design and test process in vivo. Furthermore, evolutionarily conserved codons may be recalcitrant to even synonymous mutations. Two more notable challenges await these efforts. First, even the rarest remaining codons have thousands of instances across the genome. Second, precise re-engineering of translational components will be required to discriminate between required natural codon function and reassigned codons. For example, converting a sense codon in E. coli into an open channel will require elimination of tRNA or altering the specificity of tRNA participating in wobble base-pairing interactions, and engineering of aaRSs that are capable of discriminating between natural and orthogonal tRNA.

Beyond modified inter-species barriers and the production of biomolecules with novel functions, the combination of genomic recoding and nsAA incorporation offers a novel mechanism to dynamically alter an organism’s fitness landscape. A pair of recent studies established that the recoded E. coli tolerates the insertion of nsAAs into essential genes in both functionally conserved residues (for example, the active site of a protein)¹¹⁰ and computationally redesigned protein cores¹⁰⁹. Additional genome engineering could then render the organism dependent on nsAA supplementation for viability — a condition that could not be rescued with metabolic cross-feeding or mutation in cis (FIG. 5d). These results indicate that genetic code expansion can be used to create phenotype landscapes that are dependent on the presence of synthetic amino acids.

Design principles applied to synthetic yeast chromosomes

The synthesis of the S. cerevisiae chromosome IX right arm (IXR) blended the synthesis of ~9 × 10⁴ bases with three design principles for a synthetic genome: first, an approximate wild-type phenotype; second, the removal of destabilizing elements; and third, the introduction of genetic flexibility¹⁰⁴ (FIG. 6a). By removing retrotransposons, subtelomeric repeats, introns and relocating tRNA genes, the authors sought to increase genomic stability. To increase genetic flexibility, two key changes were made. Although gene order was maintained, symmetrical loxP (loxPsym) sites¹¹⁴ were added to the 3′ untranslated regions of non-essential genes, forming the SCRaMbLE toolkit. By activating a Cre recombinase, this system enables the large-scale induction of genomic inversions and deletions, producing a large pool of genetic diversity (FIG. 6b) across a population of living cells. This mutated library of cells can then be studied for diverse emergent phenotypes in an unbiased manner. By using group II introns to insert loxP sites in microorganisms, Ellington and colleagues¹¹⁵ showed that the same recombinase system can be used to create large chromosomal inversions, insertions and deletions.

Figure 6. Design principles and assembly: synthetic yeast genome project.

a | Design principles in the synthetic yeast genome project include: wild-type phenotypes, unburdened and stabilized chromosomes via the removal of numerous types of genetic elements, and increased genetic flexibility through recoding of all TAG codons to TAA and introduction of symmetrical loxP (loxPsym) sites after non-essential genes. b | Synthetic DNA assembled in vitro is incorporated into the genome in a piece-wise manner, permitting phenotype characterization of each hierarchical segment. Induction of Cre recombinase drives activity of loxPsym, leading to chromosomal rearrangements that eliminate and translocate segments, enabling rapid exploration of dramatically altered genotypes. ORF, open reading frame; UTR, untranslated region.

The second facet of genetic flexibility was the recoding of TAG stop codons to TAA in a manner analogous to the genomically recoded E. coli^101,102. Alongside the synthetic IXR (synIXR), the authors designed a semisynthetic segment of the left arm of chromosome VI (semisynVIL). Of note, when incorporated into host strains, neither major fitness effects nor compensatory transcriptomic changes were observed. With the assistance of an undergraduate course¹¹⁶, a subsequent study carried this project to its logical conclusion with the complete synthesis of a 2 × 10⁵ bp S. cerevisiae chromosome 3 (synIII)¹⁰³. In contrast to prior work in Mycoplasma, in this first example of a synthetic eukaryotic chromosome, the conversion occurred in the target organism, rather than via an intermediate, and it remains to be seen whether this process can be replicated for the remainder of the genome. Three key distinctions of this approach are as follows: the introduction of new design features into the eukaryotic genome; an iterative synthesis and characterization approach capable of detecting deleterious mutations before final assembly; and parameters that enable future genome evolution via SCRaMbLE.

Perspectives

In his book History of Writing, Fischer suggests that “all modern society rests on writing’s plinth” (REF. 117). To date, reading capacity in the biomedical sciences has far outpaced writing capability, but recent developments reflect dramatic progress in this area. As described in this Review, numerous recent approaches for genome design and engineering have enabled a breadth of applications, including the production of biomolecules¹⁰¹, elucidating the complex genetic basis of phenotypes and strengthening the barriers between laboratory organisms and the environment^101,109,110. Near-term technological goals have been identified, including improvements in multiplexing capability³⁴ and target specificity for nuclease²⁹ and oligonucleotide methods. Furthermore, de novo synthesis and refactoring of pathways⁶⁹ will benefit from ongoing in silico improvements and integration with in vivo methods that permit combinatorial expansion of genetic landscapes⁵³, rapid functional genetic testing and characterization of new genome designs. We anticipate that breakthroughs in these areas will lead to their adoption by the wider biological community.

We envision that the future of genome engineering sits at the interface between in vivo genome evolution and de novo synthesis. As phenotypes are fundamentally unpredictable, it is essential to explore a combinatorial space of genotypes, rather than a single, static design. The scale of phenotype exploration is a factor of two variables: ease of genotypic diversification; and power of selection. De novo synthesis alone is currently capable of assembling thousands of user-defined genes, hundreds of networks and single genomes for individual iteration, but in vivo evolution strategies produce billions of combinatorial variants per day^54,57. The union of design, de novo synthesis and in vivo evolution will lead to the creation of large and diverse populations, highlighting that both computation and phenotype selections are key limitations.

The vast combinatorial complexity enabled by genome engineering technologies is too large to be explored through experiments alone. Computational approaches such as computational protein folding¹¹⁸ and flux balance analysis¹¹⁹ can be used to help to refine the space in which genome engineering is applied. Computational analyses of protein sequence or pathway conservation may help to narrow the scope of mutations to be created and evaluated¹²⁰. Additional benefits can be obtained through the use of computational genome design tools and models that are iteratively refined through multiple rounds of experiments.

Screens and selections are commonly used to evaluate mutant populations and isolate enriched mutations. Screens such as colorimetric assays and fluorescence-activated cell sorting (FACS) provide researchers with a quick and easily measured readout for the biological property under optimization¹²¹. However, screening the large number of variants created in many genome-engineering workflows with limited-throughput screening assays can quickly become impractical.

Selections provide a more efficient method for identifying high-performing variants. Given the vast space of possible mutants, this efficiency is key to identifying rare, high-performing mutants, which might be missed in a screen¹⁵. However, selective pressures can often lead to enrichment of ‘cheaters’ that have evolved an alternative path to fitness that is decoupled from the biological activity being optimized. Rounds of positive and negative selection, together with incremental increases in the stringency of selection, have been shown to reduce the prevalence of such ‘cheaters’ and improve the enrichment of high-performing variants¹²². Effective selections are dependent on establishing a strong correlation between fitness and the biological property being optimized.

Alongside screens and selections, in vivo sensors can be used to enrich complex populations of cells for desirable genotypes or phenotypes. Many metabolite-specific markers have been identified and used to optimize production in various biochemical pathways^122,123. However, the lack of in vivo sensors for all molecules of interests limits the application of this approach. Connection of protein domains that are capable of acting as biosensors to domains that are able to elicit a biological response offers promise as a method for linking biosensor recognition of a desired compound to a selectable physiological response. Biosensors that are able to link biochemical changes to a transcriptional response have been shown to be effective in metabolic engineering¹²⁴. No universal solution exists, although recent work suggests that RNA and proteins provide viable sensors for many desirable biomolecules^122,125,126.

Analogue, rather than digital, output signals offer a unique opportunity to increase the selection sensitivity for variables such as small-molecule and metabolite concentrations or gene expression levels. Colorimetric assays have already been utilized in varied genome engineering studies^53,127, but the next step is to link analogue signalling to selections in order to reduce the time and effort spent on complex phenotype characterization⁶⁹. We expect that the increasingly sophisticated circuits under development will contribute substantially to these aims¹²⁸. The development of more flexible and powerful selections will fully enable rapid genome design manifesting organisms that are capable of sampling new evolutionary and fitness landscapes.

Glossary

Directed evolution

A process that utilizes mutagenesis and a user-defined selective pressure to develop populations that are enriched for mutants exhibiting a desired function or behaviour.

Nuclease-based genome editing

The use of engineered nucleases with a DNA-targeting mechanism (for example, a zinc-finger nuclease (ZFN), transcription activator-like effector nuclease (TALEN) or a clustered regularly interspaced short palindromic repeat (CRISPR)) for targeted mutagenesis in vivo.

CRISPR–Cas

A bacterial adaptive immune system comprising clustered regularly interspaced short palindromic repeats (CRISPRs) and CRISPR-associated (Cas) endonucleases that has been repurposed for genome editing.

Recombineering

(Recombination-mediated genetic engineering). A technique that utilizes homologous recombination to alter genetic loci using synthetic single-stranded DNA and double-stranded DNA.

Bioprospecting

Searching and sourcing useful genes from other organisms to specifically confer a desired phenotype (for example, medically or industrially relevant pathways and small molecules).

Refactoring

Reorganization of biological systems or pathways with the goal of improving the ease and predictability of future engineering efforts. This often entails the removal of native regulation, separation of overlapping genetic elements and the use of well-characterized regulatory elements.

Codon optimization

A multifactorial process that utilizes synonymous codons to reflect the preferred codon frequency in order to optimize protein expression in a target species.

Gene synthesis

De novo construction of user-defined double-stranded DNA.

Genetic code expansion

Modification of the genetic code to enable site-directed incorporation of amino acids beyond the canonical 20 into proteins.

Non-standard amino acid

(nsAA). An amino acid beyond the canonical 20 present in the genetic code. nsAAs generally fall into two classes: post-translational modifications (for example, phosphoserine); and synthetic amino acids that do not exist in nature.

Orthogonal translation systems

(OTSs). Used in genetic code expansion to link incorporation of a non-standard amino acid (nsAA) to a target codon. These systems are composed of a tRNA that is selectively charged with a given nsAA by a cognate aminoacyl-tRNA synthetase (aaRS). Other possible components of OTSs include mutant elongation factor Tu and ribosomes designed to accommodate the given nsAA.

Genomically recoded organism

(GRO). An organism in which codons have been reassigned to create an alternate genetic code. For example, an Escherichia coli strain in which all instances of the TAG stop codon have been mutated to TAA and release factor 1 (RF1) has been deleted, enabling the use of TAG as an open coding channel.