|Home | About | Journals | Submit | Contact Us | Français|
Genome-modification technologies enable the rational engineering and perturbation of biological systems. Historically, these methods have been limited to gene insertions or mutations at random or at a few pre-defined locations across the genome. The handful of methods capable of targeted gene editing suffered from low efficiencies, significant labor costs, or both. Recent advances have dramatically expanded our ability to engineer cells in a directed and combinatorial manner. Here, we review current technologies and methodologies for genome-scale engineering, discuss the prospects for extending efficient genome modification to new hosts, and explore the implications of continued advances toward the development of flexibly programmable chasses, novel biochemistries, and safer organismal and ecological engineering.
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 16–17bp 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.
Because delivering large genetic constructs into many cell types is difficult, highly efficient methods of recombining the host genome with an introduced construct are useful for applications requiring large amounts of foreign DNA or the replacement of many contiguous genes with modified or synthetic variants. Recombinases are DNA-binding enzymes that catalyze highly specific and efficient DNA splicing reactions between two sites. Early experiments with phage-derived recombinases irreversibly incorporated circular constructs containing the phage attP site into the attB site of the host genome normally utilized by the phage (Mizuuchi and Mizuuchi, 1980). Later work demonstrated that these ‘integrases' can perform a similar role in a wide variety of species if the appropriate attB or attP target site is inserted into the genome by other means (Kilby et al, 1993), or may alternatively utilize ‘pseudo-att' sites native to the genome at somewhat lower efficiency (Thyagarajan et al, 2001). Cre recombinase, originally from phage P1 (Sternberg et al, 1981), is the gold standard for efficient recombination of target sites across a wide variety of species. However, its comparative promiscuity leads to toxicity in some eukaryotes, leading to the development of Flp recombinase as an alternative (Turan et al, 2011). Unlike integrases, Cre and Flp are reversible enzymes that normally recombine two identical recognition sites to invert or excise the intervening sequence, but they can be made irreversible by utilizing ‘poisoned' half-sites that generate an inactive site upon recombination (Schlake and Bode, 1994; Albert et al, 1995).
In the context of genome-scale engineering, recombinases are most useful for efficiently inserting large DNA constructs into the genome. By flanking an endogenous sequence with orthogonal recognition sites from two different recombinases or two orthogonal sites recognized by the same recombinase, the sequence may be replaced by a synthetic donor construct containing compatible sites (Schlake and Bode, 1994; Missirlis et al, 2006; Sheren et al, 2007). With three pairs of orthogonal sites, this technique could conceivably be used to iteratively insert large cassettes into the genomes of many different organisms ad infinitum (Turan et al, 2011; Obayashi et al, 2012). Unfortunately, recombinases require pre-existing recognition sites, which must be introduced to the target site by another method. Although directed evolution methods have yielded recombinases capable of recognizing alternative sites (Buchholz and Stewart, 2001; Sarkar et al, 2007), such approaches are presently too laborious for most laboratories. New methods of performing directed evolution may relax this limitation (Esvelt et al, 2011). A promising design-based alternative involves replacing the native DNA-binding domain with an exogenous domain that can be more easily engineered to target a sequence of interest (Akopian et al, 2003). Although the resulting chimeric enzymes are highly specific, they are currently inefficient compared with natural recombinases (Gordley et al, 2009). It is likely that extensive directed evolution will be required to render the catalytic domain suitable for retargeting by replacement of the DNA-binding domain.
Targeted genome engineering requires a means of specifically recognizing the sequence of each site to be modified. Zinc-fingers (ZFs) and TAL (transcription activator-like) effectors are a class of versatile and programmable DNA-binding proteins that have enabled effector proteins, including DNA-modifying enzymes, to be targeted to specific sequences in a variety of organisms. ZFs are stackable motifs of ~30 amino acids that recognize approximately three base pairs of DNA with varying specificity. Although ZFs recognizing each triplet cannot be simply stacked to reliably recognize longer sequences (Ramirez et al, 2008), a variety of design (Sander et al, 2011b) and selection-based (Maeder et al, 2009) methods are capable of generating specific DNA binders. Unfortunately, custom ZFs remain relatively difficult and expensive to obtain for the typical laboratory. DNA recognition by TAL effector domains is more straightforward, with each 34-aa TAL motif recognizing a single basepair through contacts with amino acids 12 and 13, known as the repeat variable di-residue (RVD) (Boch et al, 2009). Unlike ZFs, TAL effectors are readily stacked to recognize long sequences. Although the assembly of TALs is complicated by their larger size and abundant repeat regions, a number of recently described approaches have the potential to overcome these challenges (Weber et al, 2011; Briggs et al, 2012; Reyon et al, 2012).
ZF and TAL nucleases (ZFNs and TALENs) are created by coupling a ZF or TAL DNA-binding domain to the nonspecific nuclease domain of the FokI restriction enzyme. When two monomers bind to adjacent sites, their FokI domains dimerize and catalyze DNA cleavage, causing a double-strand break (DSB) (Kim et al, 1996). DSBs are most commonly repaired by homologous recombination (HR) or non-homologous end-joining (NHEJ). ZFN cleavage followed by HR with a donor sequence containing homologous flanking regions leads to insertion of the donor sequence at efficiencies of ~1–15% (Urnov et al, 2005), while ZFN cleavage followed by error-prone NHEJ results in gene disruption from small deletions or insertions, typically at somewhat higher efficiencies (Urnov et al, 2010). Targeted gene editing using ZFNs has been demonstrated in a variety of cell types, including flies (Bibikova et al, 2003), worms (Wood et al, 2011), sea urchins (Ochiai et al, 2010), zebrafish (Ekker, 2008), silkworms (Takasu et al, 2010), frogs (Young et al, 2011), plants (Cai et al, 2009; Osakabe et al, 2010; Zhang et al, 2010), and numerous mammals (Urnov et al, 2005; Geurts et al, 2009; Hauschild et al, 2011). Nuclease activity can be toxic in some cell types, possibly due to off-target activity, but this problem can be mitigated by utilizing less toxic ‘nickase' variants that cut only one strand (Kim et al, 2012; Ramirez et al, 2012). Customized ZFNs are commercially available, although at a significant cost. TAL effector nucleases (TALENs) can more readily target a variety of sequences by virtue of their more flexible RVD-based recognition. Although newer and less thoroughly studied, TALENs appear to have fewer off-target effects and lower toxicity than corresponding ZFNs (Mussolino et al, 2011). Design tools are freely available (Doyle et al, 2012) with predicted viable cleavage sites every 35 basepairs in mammalian genomes on an average (Cermak et al, 2011). Their primary weakness is the difficulty of assembling and delivering such large and repeat-prone sequences. TALENs have been successfully applied in numerous organisms including yeast (Li et al, 2011), flies (Liu et al, 2012), zebrafish (Sander et al, 2011a), plants (Li et al, 2012), rats (Tesson et al, 2011), and human cells (Hockemeyer et al, 2011) with gene disruption efficiencies of up to 25% (Miller et al, 2011).
Certain group II introns are selfish genetic elements that undergo genomic transposition through an RNA intermediate. Because targeting is determined primarily by base-pairing interactions with the intron RNA, these site-specific retrotransposons can be retargeted to accomplish both gene disruption and gene insertion. The commercially available Targetron system harnesses a retrotransposon capable of inserting up to 1.8kb into the genome (Karberg et al, 2001). Intron retrotransposition efficiencies vary from 1–80% depending on the site and species (Perutka et al, 2004). Sequences suitable for insertion are found every few hundred bases on average, permitting most genes to be disrupted. Moreover, the system is active in a wide variety of microbes, providing genetic manipulation of species that cannot be modified using other methods (Yao and Lambowitz, 2007). Notably, insertions of recombinase recognition sites may permit subsequent recombinase-mediated cassette exchange. Targeting efficiency may be high enough to permit multiplex modifications, though this has yet to be demonstrated. Interestingly, group II introns can also be used to generate DSBs (Karberg et al, 2001), suggesting a potential use in promoting HR if they can be engineered or evolved to function efficiently in eukaryotes.
Recombineering (or recombinogenic-engineering) uses a phage-derived HR pathway to recombine a donor DNA strand with a homologous sequence in the bacterial host. Given sufficient regions of flanking homology (>500bp), endogenous HR, which is usually mediated by the RecA/Rad51 pathway, is capable of integrating sequences into the genome of almost any cell. However, low efficiency of the native HR machinery limits the use of this technique without efficient DNA delivery and selection. Recombineering is an improved approach that utilizes phage proteins (RecET, λ-Red) to dramatically increase HR frequencies across the entire genome (Zhang et al, 1998; Datsenko and Wanner, 2000; Yu et al, 2000). In E. coli, HR by λ-Red is RecA-independent and instead relies on three proteins: Exo, Beta, and Gam (Muyrers et al, 2000; Yu et al, 2000). Exo is a 5′→3′ exonuclease that digests linear double-stranded DNA (dsDNA), leaving 3′ single-stranded intermediates that then act as substrates for subsequent recombination (Maresca et al, 2010). Beta is a single-stranded DNA (ssDNA)-binding protein that facilitates recombination via hybridization of the linear fragment to its genomic complement. Gam acts to inhibit RecBCD activity in vivo to prevent the degradation of foreign linear dsDNA fragments. Although recombineering still requires a selection step, the λ-Red-like system will function with as few as 40bp of homology flanking double-stranded donor DNA fragments of up to several kilobases in length, a limit imposed by a combination of the transformation and recombination efficiencies. Thus, simple PCR amplification of a selectable cassette (typically an antibiotic resistance or metabolic gene), with primers containing flanking homologous sequences to the target site, enables limited rewriting of any region of the genome (Sharan et al, 2009). A recent combinatorial example of this technique, Trackable Multiplex Recombineering, used primers derived from DNA microarrays to generate pools of barcoded dsDNA cassettes that can target different sites across the genome (Warner et al, 2010). Short ssDNA can also be used in recombineering, a process which requires only the λ-Beta protein. We discuss the utility of such approaches for multiplexed recombineering in the next section. Although recombineering systems have been developed for several model bacteria (van Kessel and Hatfull, 2007; Swingle et al, 2010a; van Pijkeren and Britton, 2012), more work is needed to expand the methodology to other organisms. A search for λ-Red-like enzymes derived from phages and viruses that infect other organisms is ongoing (Datta et al, 2008).
The nucleic acid-targeted CRISPR (Clustered Regularly Interspaced Short Palindromic Repeat) system has great potential for genome modification in many organisms. CRISPR systems defend bacteria and archaea from invading phage and plasmids by RNA-directed degradation of DNA (Wiedenheft et al, 2012). In Type II CRISPR systems, the Cas9 protein locates DNA ‘protospacer' sequences homologous to the ‘spacer' sequence in a guiding CRISPR RNA (crRNA) and checking for sufficient RNA–DNA base pairing (Jinek et al, 2012). Upon identifying a matching sequence that also contains an appropriate protospacer-adjacent motif (PAM), the enzyme cleaves both DNA strands ~3bp from the start of the PAM, causing a DSB (Gasiunas et al, 2012). PAM sequences are quite short (NGG (Deltcheva et al, 2011), NGGNG (Horvath et al, 2008), NNAGAAW (Deveau et al, 2008), and NAAR (van der Ploeg, 2009) to date), permitting most sequences to be targeted. At least 12bp of perfect homology, in addition to the PAM, appears to be necessary for CRISPR endonuclease activity (Deveau et al, 2008; Sapranauskas et al, 2011; Jinek et al, 2012; Mali et al, 2013; Cong et al, 2013). In bacterial CRISPR loci, the spacer regions of crRNAs are normally flanked by direct repeats of similar size that are critical for recognition and processing by Cas9 and RNaseIII (Deltcheva et al, 2011), but synthetic mimics of the mature crRNA function equally well in vitro (Jinek et al, 2012).
We and others have recently demonstrated that Cas9 can be used to engineer mammalian genomes (Mali et al, 2013; Cong et al, 2013). Cas9 can be directed to cleave any sequence with a compatible PAM—in these cases NGG—by expressing a chimeric RNA mimic (Mali et al, 2013) or a spacer array together with the tracrRNA required for processing (Cong et al, 2013). Gene modification via DSB-stimulated HR is accomplished by simply expressing Cas9 and a cassette that generates a RNA with a spacer matching the target sequence in the desired cell. Targeting two adjacent sites effectively deleted the intervening region, demonstrating limited but multiplexed gene disruption capabilities. Knocking out one of the two Cas9 nuclease domains converted the enzyme into a nickase capable of stimulating HR with comparable efficiency while reducing the frequency of NHEJ. Importantly, both gene disruption and HR rates appear to be comparable to or greater than those achieved with ZFNs and TALENs targeting the same loci.
Interestingly, sustained Cas9 activity might be used to simultaneously promote HR while selecting against cells retaining the target region, potentially obviating the need for positive selection markers. This approach would be feasible in genomes engineered to constitutively express Cas9, which could be subsequently edited by simply delivering the appropriate donor cassette and crRNA. Further development of CRISPR-mediated genome engineering technologies should focus on increasing the specificity beyond the current 12bp+NGG sequence, which would likely lead to some unintended off-target cutting, and on enabling genomic sequences with alternative PAMs to be targeted. Due to its significantly greater ease of use, Cas9-mediated gene targeting represents a new and promising genome editing approach, especially in mammalian systems.
The ability to edit single genes is an important step toward engineering whole genomes. The explosion of modifications achieved with ZFNs and TALENs are particularly striking given the dearth of prior alternatives for most multicellular organisms. Still, the sheer size of even the smallest bacterial genomes renders serial modification of limited utility for truly genome-scale engineering endeavors. Efficient methods enabling multiplex genome editing are urgently needed.
Unfortunately, techniques that generate DSBs to catalyze homology-directed repair may be difficult to multiplex due to the toxicity of multiple simultaneous breaks and the high rate of NHEJ, which could easily lead to unintended rearrangements. High-efficiency ZF or TAL effector recombinases represent one potential alternative, although quickly generating large numbers of ZFs or TALs presents an additional challenge. Another option might involve fusing a nuclease-inactivated Cas9 protein to the catalytic domain of a recombinase, although retaining function could prove to be difficult. Group II introns may be multiplexable for gene disruption, but they leave unavoidable scar sites, are limited to small cargo capacities, and have not been demonstrated to work efficiently in eukaryotes. Meanwhile, the low efficiency of double-stranded λ-Red-mediated recombineering also limits its use for multiplexed genome-scale engineering. However, λ-Red-like proteins also facilitate recombination of smaller ssDNA fragments. On the basis of prior work (Ellis et al, 2001), we recently described an approached known as Multiplex Automated Genome Engineering (MAGE) that utilizes short ssDNA oligonucleotides (oligos) instead of dsDNA cassettes to mediate targeted genome modification (Wang and Church, 2011; Wang et al, 2009). Specifically, oligos that are complementary to the lagging strand of replicating genomes are incorporated into the daughter genome at high efficiency, presumably by mimicking Okazaki fragments at the replication fork (Yu et al, 2003). Oligos that target the leading strand appear to have >50-fold lower incorporation efficiency.
MAGE can precisely engineer any site in the genome by simply introducing an oligo matching the desired sequence. Oligos ranging from 30 to 100 bases are efficiently integrated as long as there are sufficient homology arms to facilitate ssDNA annealing to the target (Ellis et al, 2001). At the center of the oligo, new sequences can be designed (up to 30 bases along a 90-base oligo) and introduced into the genome as a heteroduplex, which is resolved into fully mutated alleles during subsequent rounds of cell division. In E. coli, oligo incorporation is increased >1000-fold by the ssDNA-binding protein λ-Beta. Removal of the endogenous mismatch repair machinery (e.g., ΔmutS) (Costantino and Court, 2003) or evasion of mismatch repair through modified bases (Wang et al, 2011) can significantly increase the efficiency of oligo incorporation to levels >30% per viable progeny (Wang et al, 2009). Use of a co-selectable marker can further increase the efficiency to >70% (Carr et al, 2012; Wang et al, 2012b).
Several factors make the oligo-mediated MAGE approach particularly attractive for genome-scale engineering. First, the transformation efficiency of short oligos is high compared with plasmids or dsDNA cassettes, thereby allowing large pools of oligos with different genomic targets to simultaneously enter the cell and undergo incorporation. Because not all oligos are incorporated in every cell, combinations of mutations are generated through this process. With incorporation efficiencies above 70%, cells containing >10 targeted mutations can be isolated after a single transformation (Lajoie et al, 2012) by simply screening 100 colonies with multiplex allele-specific PCR (Wang and Church, 2011). Second, the protocol can be iteratively repeated on a population of cells with only 2–3h of recovery growth needed between cycles. Iterative cycling enables further multiplexing and enrichment of mutants that are otherwise found at low frequencies in the population, which can be automated (Wang et al, 2009). Third, oligos can be easily and cheaply synthesized using commercial vendors and used directly in MAGE reactions without the need for further processing, in contrast to dsDNA cassettes which require additional steps of PCR amplification and purification. Furthermore, high-density DNA microarrays can serve as potential sources of large pools of unique DNA sequences to extend multiplexed genome-scale engineering. Finally, oligo-mediated genome engineering approaches such as MAGE will likely function in a variety of organisms by virtue of mechanistic simplicity. To date, oligo-mediated allelic replacement has been demonstrated in Gram-negative bacteria (Swingle et al, 2010b), Gram-positive bacteria (van Pijkeren and Britton, 2012), and mammalian cells (Rios et al, 2012).
Since the chemical synthesis of the first gene in 1972 (Agarwal et al, 1972), the cost of DNA synthesis has precipitously decreased as the throughput has soared, enabling construction and assembly of genes and genomes de novo (Carr and Church, 2009). Individual gene-sized DNA fragments are readily synthesized commercially and assembled into larger operons (Kodumal et al, 2004; Tian et al, 2009). Efforts to build phage (Chan et al, 2005) and viral genomes (Blight et al, 2000; Cello et al, 2002), chromosomal arms of S. cerevisiae (Dymond et al, 2011), and, most impressively, the entire genome of M. mycoides (Gibson et al, 2008) have been described. New technologies enabling oligonucleotide synthesis on DNA microarrays continue to reduce the cost and increase the throughput for building synthetic genes and genomes (Tian et al, 2004; Kosuri et al, 2010; Quan et al, 2011).
The question of when it is best to adopt an editing, semi-synthetic, or synthetic approach to genome engineering hinges on the reliability of design. Without the ability to accurately evaluate large numbers of potential designs in silico, we must build and test them empirically. Currently, large-scale de novo synthesis of a genome requires a significantly greater level of resources and effort than directly editing an existing genome. Consequently, a genome editing approach may be optimal when generating genomes with a moderate degree of specified changes (i.e., ≤100s of changes, <100bp each), as is required for tuning regulatory networks (Wang et al, 2009, 2012b) or altering protein sequences (Wang et al, 2012a). A de novo synthesis approach is more likely to be appropriate for larger-scale alterations such as codon optimization (Welch et al, 2009) or refactoring (Chan et al, 2005; Temme et al, 2012) that are recalcitrant to genome editing technologies.
Building an entire synthetic genome can be difficult to troubleshoot, costly, and prone to failure. An illustrative example of such issues was observed during the construction of the synthetic 1.1Mb M. mycoides genome, when a single basepair deletion in the essential gene dnaA prevented the generation of a viable cell (Gibson et al, 2010). Only when different synthetic pieces were swapped with natural sequences did the researchers identify the source of the error, highlighting the importance of direct testing. Underlying design flaws may be even more difficult to assess as they may impact the cell physiology in non-linear and epistatic ways. Thus, step-wise construction and testing of progressively modified intermediates will be a crucial approach for most genome-scale engineering efforts until the failure rate of engineered biological designs can be reduced to acceptable levels. Consequently, methods capable of rapidly assembling and exchanging individually synthesized and separately tested genome fragments will be needed. Current examples include in vitro enzymatic assembly methods (Li and Elledge, 2007; Engler et al, 2008; Gibson et al, 2009; Zhang et al, 2012) and Conjugative Assembly Genome Engineering in vivo (Isaacs et al, 2011) (Figure 4). Recent studies have already described instances of cloned or hybrid genomes constructed by transformation or assembly of a donor genome into a recipient cell that retains its own genome. While the Bacillus-Synechocystis hybrid-genome (Itaya et al, 2005) and the S. cerevisiae clone containing a copy of the A. laidlawii genome (Karas et al, 2012) have yet to yield useful new phenotypes, they do illustrate cellular robustness to large-scale genomic insertions. Studies that evaluate the effects of swapping or refactoring essential operons will provide information more directly relevant to evaluating the feasibility of new designs. More generally, developments that further combine synthetic, semi-synthetic, and hybrid approaches will lead to deeper understanding of the limits of rational design and optimization for engineered biological systems.
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.
The difficulties inherent in designing living systems arise from the vast number of cellular components and the sheer complexity of their evolutionarily optimized network of interactions. Simulating large numbers of heterogeneously interacting molecules requires evaluating the probability and magnitude of all possible interactions between non-identical components, a task that would be computationally beyond us even if we had perfect knowledge of every interaction (Koch, 2012). We still do not understand the function of almost 20% of the ~4000 genes found in E. coli (Keseler et al, 2011). Given that biological complexity is one of the most significant barriers to rational genome design, we should aim to build a simplified microbial cell. Not only would such a cell serve as an improved chassis for future engineering, the act of constructing such a genome will transform our understanding of the factors contributing to the performance, evolvability, and robustness of cellular systems in general.
Single-gene deletion experiments (Giaever et al, 2002) suggest that a significant number of all genes are redundant, with only ~300 being individually essential (Feher et al, 2007). The first step toward a simplified cellular chassis is to reduce the genome to a functionally useful set of genes. Several groups have embarked upon endeavors to eliminate all nonessential genes, starting with E. coli (Hashimoto et al, 2005; Posfai et al, 2006), B. subtilis (Ara et al, 2007), and S. pombe (Giga-Hama et al, 2007). It is important to keep in mind that whether a gene is essential depends on the environmental conditions. Therefore, we define a set of useful traits for a biological chassis as (1) fast growing in minimal media with glucose, (2) capable of fermentation, (3) amenable to genetic manipulation, and (4) minimally sufficient such that removal of any additional gene negatively affects the other three stated considerations. A cell containing a set of genes that satisfy the above criteria is said to have a core or minimal chassis. Although a viable E. coli genome with 20% fewer genes has already been engineered (Posfai et al, 2006), it is likely that a reduction of 50% is achievable for the core chassis. Even though smaller genomes and simpler transcriptome do exist (e.g., Mycoplasma pneumonia (Guell et al, 2009)), our core chassis will be much more useful for biological engineering because it will not suffer from slow growth or depend upon additional exogenous metabolites. Moreover, engineering our chassis could consolidate related genes into modular, functionally similar operons to facilitate future engineering.
With far fewer components and exponentially fewer possible interactions, a cell with a core chassis will be much more amenable to in silico modeling than wild-type E. coli or even M. genitalium (Karr et al, 2012). Still, its remaining components will interact in many more ways than we would prefer, and not all of them are understood. This might be remedied by reducing the number of regulatory interactions, ideally by replacing endogenous regulatory elements with well-defined orthogonal equivalents. Temme et al (2012) implemented this concept by ‘refactoring' the nitrogen fixation cluster to remove all native gene regulation. Refactoring an operon involves removing all non-coding DNA, nonessential genes, and transcription factors, replacing essential genes with computer-designed synthetic genes recoded to eliminate internal regulatory sites, and adding synthetic regulation. Extending this approach to the entire core genome will be an immense challenge, as each replacement must be optimized with synthetic components. On the other hand, cellular growth and survival is a powerful and readily applicable selection, enabling libraries of synthetic or rewired regulatory elements to be quickly selected and sequenced to identify the best performers (Isalan et al, 2008). Minimizing the total number of orthogonal regulatory elements and compensating for changes in the expression of previously refactored operons caused by adding additional binding sites are likely to be the most challenging aspects of the project. Adding additional but well-defined levels of regulation such as orthogonal 16S ribosomes (Rackham and Chin, 2005), synthetic ZF transcription factors (Khalil et al, 2012), or orthogonal RNA-based translational repressors (Isaacs et al, 2004) may be necessary to increase growth to acceptable levels while minimizing the total number of components.
A final challenge concerns the effects of natural selection on our simplified genome. We expect our rationally designed synthetic chassis to be suboptimal, in that simple growth in glucose media may lead to accumulation of beneficial mutations. Careful tracking of these beneficial mutations as they occur will simplify the task of decoding the newly created interactions and reveal important design flaws in our in silico models. Only by understanding and attempting to compensate for these new interactions will we learn how to further simplify and optimize the performance of our engineered system.
A frequent objection to the use of genetically modified organisms is the possibility of unintended consequences arising from accidental release. Improved methods for biological containment would reduce such risks while raising public awareness of beneficial genome engineering research. One such containment strategy is the development of a chassis that utilizes an orthogonal genetic code (Isaacs et al, 2011). The canonical encoding scheme maps 64 possible codons to 20 corresponding amino acids and three stop signals. Except for a few known organisms (Knight et al, 2001), the genetic code is the single most well-preserved property in all of biology and thought to be irreversibly fixed in its current configuration as a result of ‘the frozen accident' (Crick, 1968). A codon-swapped organism might have codons that are normally assigned to leucine instead encode arginine. Although the resulting protein sequence would not change, the encoded nucleotide sequences would be quite different in a recoded organism compared with the wild type. Achieving this goal would involve not only recoding of all genes in the new genomic chassis, but would also require minor alterations to the anticodon sequences of tRNAs to accommodate different codon swaps. A combination of genome synthesis and engineering will be needed to realize such an endeavor.
More importantly, a radically recoded chassis would be unable to productively exchange genetic material with other organisms in the environment. When transferred into a wild-type cell, recoded genes from a swapped-codon chassis will generate meaningless proteins due to mistranslation from reassigned codons. Conversely, natural genes will not function in the swapped-codon chassis, preventing our synthetic genome from becoming contaminated with wild toxins, pathogenicity elements, or antibiotic resistance genes. Indeed, genetic isolation from all other domains of life will also confer broad immunity to natural viruses, a significant advantage for the industrial-scale production of biochemicals. However, the recoded chassis may still interact with the physical environment and with other organisms indirectly via nutritional exchange and space competition. These aspects present opportunities for further rational engineering. Finally, recoded organisms will contain many genomic signatures of their synthetic origin, allowing easy identification and surveillance of their origin, make, and purpose in comparison to natural variants.
With the exception of post-translational modifications in higher-level organisms, the amino-acid repertoire of cells is mostly confined to the canonical 20 amino acids. Unnatural amino acids have been successfully incorporated into proteins using several strategies involving orthogonally evolved tRNA and tRNA synthetases (Hendrickson et al, 2004; Xie and Schultz, 2005), but this approach has been hampered by lower efficiencies of incorporation due to competition with existing codon recognition factors (Young et al, 2010). Expanding the repertoire of possible amino acids that the cell can use to build proteins is a powerful capability that will be readily available to any recoded chassis. Unnatural amino acids will dramatically expand the biochemical repertoire of cells by enabling new chemistries that are inaccessible to natural systems (Liu and Schultz, 2010). Whole-genome recoding can readily free up codons by reducing the degeneracy of the current codon mapping. New amino acids can be assigned to ‘free codons' as long as the existing proteins are recoded with the synonymous codons to retain the amino-acid sequence. A similar event occurred when a handful of organisms began to encode the 21st amino acid, selenocysteine, with the TGA codon that functions as a STOP codon in other forms of life (Forchhammer et al, 1989). Although eliminating significant numbers of rare sense codons may be challenging, the prospect of engineering a flexible chassis with the ability to encode multiple unnatural amino acids and access phenotypes unavailable to natural organisms is worth the attempt.
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.
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.
The authors declare that they have no conflict of interest.