Rapid Prototyping of Microbial Cell Factories via Genome-scale Engineering

Abstract

Advances in reading, writing and editing genetic materials have greatly expanded our ability to reprogram biological systems at the resolution of a single nucleotide and on the scale of a whole genome. Such capacity has greatly accelerated the cycles of design, build and test to engineer microbes for efficient synthesis of fuels, chemicals and drugs. In this review, we summarize the emerging technologies that have been applied, or are potentially useful for genome-scale engineering in microbial systems. We will focus on the development of high-throughput methodologies, which may accelerate the prototyping of microbial cell factories.

1. Introduction

Microbial cell factories (MCFs), which convert biomass resources to value-added compounds such as fuels, chemicals, materials and pharmaceuticals, have been proposed as a sustainable and renewable alternative to the traditional petrochemical-based processes (Keasling, 2010, Lee et al., 2012, Rabinovitch-Deere et al., 2013). However, intensive reprogramming of cellular metabolism is required to achieve economically feasible fermentation processes with MCFs. Conventional strain engineering approaches rely on random mutagenesis, which is achieved through chemical mutagens/UV irradiation (Crook and Alper, 2012), prolonged cultivation under selective pressure (Portnoy et al., 2011), transposon insertions (Eckert et al., 2011, Hamer et al., 2001, Hutchison et al., 1999) and genome shuffling (Biot-Pelletier and Martin, 2014, Zhang et al., 2002). Effective in generating improved phenotypes using simple techniques, these methods are widely adopted in industry, especially for those host organisms with poorly defined genetics and limited engineering tools (Crook and Alper, 2012). However, traditional approaches are often labor-intensive, time-consuming, and difficult to analyze and transfer the genetic basis of a selected trait. Recently, the scale, efficiency and precision of genetic analysis and manipulation have been remarkably improved by several enabling technologies, including but not limited to microarray DNA synthesis, next-generation DNA sequencing (NGS), programmable DNA-binding proteins, and in vivo biosensors. Nowadays, billions of genome variants can be created in a directed and/or combinatorial manner, and the mutant strains with the optimal performance can be rapidly isolated. Collectively, these new technologies and their applications exemplify an emerging discipline called ‘genome engineering’ or ‘genome-scale engineering’ (Carr and Church, 2009, Esvelt and Wang, 2013, Jeong et al., 2013, Segal and Meckler, 2013).

The practice of genome-scale engineering can be broadly classified into three categories: genome editing, transcriptome engineering, and genome synthesis. Genome editing precisely or combinatorially modifies the target genome at multiple loci. Modifications are located either in the open-reading frames (ORFs) or in the cis-acting regulatory elements such as promoters and ribosome-binding sites (RBSs). Transcriptome engineering essentially targets trans-acting regulatory elements, such as transcription factors (TFs) or non-coding RNAs (ncRNAs), by mutating endogenous regulators or introducing artificial ones. Genome synthesis involves hierarchical assembly of short chemically synthesized DNA fragments into viral/microbial genomes and yeast chromosomes. Although current synthetic genomes are constructed mainly based on their wild type templates, the ultimate goal is to write genome sequences de novo.

In this review, we first introduce the recent development in genome editing (section 2), transcriptome engineering (section 3), and genome synthesis (section 4). We then highlight how these techniques can facilitate high-throughput genotyping and phenotyping (section 5), which greatly accelerates our understanding and engineering of microbial genomes. In addition, we will discuss several examples on the application of genome-scale engineering to improve MCF performance and provide perspectives on how computational approaches and laboratory automation can be further integrated.

2 Genome editing

Unlike random mutagenesis, targeted genome editing results in elaborative and massive genome modifications with a traceable manner. Homologous recombination (HR) is the core mechanism of most targeted genome editing techniques, and various enzymes have therefore been investigated to either mediate or promote HR in microorganisms.

2.1 Recombinases

Recombinases catalyze exchange of short homologous regions (30~40 bp) of DNA. Site-specific recombinases are grouped into two families, the tyrosine recombinase family and the serine recombinase family (Turan et al., 2013). An early characterized member of the tyrosine recombinase family was λ integrase, which enables incorporation of phage DNA into the bacterial chromosomes. The λ integrase mediates irreversible recombination between the attP and attB sites in the phage and host chromosomes respectively, generating recombinant attL and attR sites (Mizuuchi and Mizuuchi, 1980). Later, Cre (from phage P1) and flippase (FLP, from the 2μ plasmid of yeast Saccharomyces cerevisiae), recognizing the loxP site and the flippase recognition target (FRT) site, respectively, were widely used for efficient recombination in a variety of species (Dymecki, 1996, Nagy, 2000, Sternberg et al., 1981, Turan et al., 2011). With identical recognition sites, Cre and FLP can reversibly invert, integrate or excise DNA sequences between recognition regions. Alternatively, such processes can be made irreversible using a partially mutated recognition site to yield a poorly recognized region after recombination (Albert et al., 1995, Schlake and Bode, 1994). For the serine recombinase family, ϕC31 integrase (from Streptomyces phage ϕC31) was the most well-studied example. Behaving like the λ integrase, ϕC31 was proven to have great potential in eukaryotic genome engineering (Karow and Calos, 2011).

In addition to inversion, integration and excision facilitated by the above-mentioned recombinases, recombinase-mediated cassette exchange (RMCE) is another useful approach in genome engineering. By flanking the target genomic locus with two different spacer mutant (“heterospecific”) sites recognized by the same recombinase or orthogonal sites of different recombinases, the endogenous region will be replaced by a donor cassette with compatible recognition sites (Turan, Zehe, 2013). Many Cre and FLP variants were engineered to recognize different sites with little cross reactivity with the wild type system, allowing efficient directional cassette exchange (Fig. 1A) (Buchholz and Stewart, 2001, Schlake and Bode, 1994, Turan et al., 2010). By exploiting the specific attP×attB recombination event, ϕC31 was also applied to cassette exchange without the requirement of heterospecific att-sites. However, ϕC31 mediated-cassette change was in a unidirectional manner (Turan and Bode, 2011).

Fig. 1. Overview of various genome-scale engineering tools.

(A) A circular donor cassette can be integrated into the recognition site by an integrase. By flanking a target sequence with heterospecific sites, RMCE enables replacement of a target sequence using a donor cassette flanked by compatible sites. (B) The ss-oligos containing designed mutations are incorporated into the lagging strand of replicating DNA through recombineering. (C) HR or NHEJ is greatly promoted via DSBs using various endonucleases. (D) Site specific insertion can be achieved via group II introns. (E) Transcription factor (TF) libraries can be constructed by mutating the endogenous TFs (gTME) or introducing artificial TFs for large-scale perturbation on transcriptome. (F) Different regulatory non-coding RNAs (ncRNAs), including sRNAs, siRNAs and gRNAs, are used to modulate targeted gene expression in bacteria and yeast.

Notably, pre-existing recognition sites are required for all events mediated by recombinases. Therefore, introduction of recognition sites into the target locus is unavoidable, which limits the application of recombinase-based methods for genome editing. Although much effort has been invested in the directed evolution of recombinases with new target recognition sequences, the engineered enzymes were inefficient in most cases (Gordley et al., 2009).

2.2 Recombination-mediated genetic engineering (Recombineering)

Taking advantage of bacteriophage-based recombination proteins (e.g., Red-Exo, Beta and Gam from λ phage (Murphy, 1998) and Rec E/T from Rac prophage (Zhang et al., 1998)), recombineering enables efficient and large-scale recombination between the transformed DNA fragments and the bacterial genome (Sharan et al., 2009). Recombination through the λ Red system relies on three proteins: Exo (also known as α), Beta and Gam. Exo is a 5’to 3’ exonuclease that digests double-stranded DNA (dsDNA) with 3’ overhangs generated. Beta is a single-stranded DNA (ssDNA) binding protein, which facilitates recombination between target locus and donor DNA (Fig. 1B). Gam inhibits endogenous RecBCD and SbcCD activity, preventing degradation of exogenous DNA by the host (Datta et al., 2008). In the RecE/T system, RecE is a 5’to 3’ exonuclease (similar to Exo) and RecT is a ssDNA-binding protein that stabilizes ssDNA and promotes recombination (similar to Beta). Either dsDNA or ssDNA can be employed for recombineering (Ellis et al., 2001, Sawitzke et al., 2007). With a linear donor DNA flanked by homology sequences as short as 40 bp, efficient recombination can be achieved by the λ red system (Sharan, Thomason, 2009). In addition, the RecET proteins are more efficient than the λ red proteins, especially in case of HR between two linear molecules (Fu et al., 2012). When these bacteriophage proteins were expressed in E. coli, the recombination efficiencies dramatically increased from 1 per 10⁶ cells to 1 per 10^3–4 cells without optimization (Boyle et al., 2013, Murphy, 1998, Swingle et al., 2010, Zhang, Buchholz, 1998).

With further optimization (e.g., modifying the cellular machinery involved in DNA replication (Lajoie et al., 2012) or optimizing the concentration and length of donor oligos (Sawitzke et al., 2011)), recombineering technology has been exploited for multiplexed genome engineering. Notably, multiplex automated genome engineering (MAGE) and trackable multiplex recombineering (TRMR) accelerated E. coli genome evolution and analysis by creating genome-wide combinatorial genetic modifications in a short time (Wang et al., 2009, Warner et al., 2010). These studies are discussed in greater details in later sections.

2.3 Endonucleases

Apart from recombineering, introduction of a double-strand break (DSB) to a chosen site is another important strategy for targeted genome editing. Normally, a DSB is repaired by either HR or non-homologous end joining (NHEJ). Given a donor sequence flanked by homologous regions to target locus, HR can be facilitated by a DSB (Iglehart and Silver, 2009). In some hosts with low HR efficiency (e.g., mammalian cells), error-prone NHEJ is boosted by DSBs, which results in gene disruption (Sun et al., 2012b). Taken together, DSB-promoted HR or NHEJ eases the arduous targeted genome editing in most organisms. As a result, different natural and artificial endonucleases have been developed and engineered to introduce DSBs at desired loci (Fig. 1C).

2.3.1 I-SceI meganucleases

I-SceI is a homing endonuclease from the mitochondria of Saccharomyces cerevisiae and was the first endonuclease used in genome engineering (Silva et al., 2011). Compared to traditional restriction enzymes, I-SceI recognizes a longer sequence (18 bp), rendering it more rare-cutting (once in every 6.9×10¹⁰ bp). Indeed, HR promoted by I-SceI has been demonstrated in several prokaryotes and eukaryotes (Cox et al., 2007, Maggert et al., 2008, Yu et al., 2008). However, like recombinases, the requirement to introduce a recognition site into a target locus limits the application of I-SceI-based methods. Directed evolution of I-SceI towards a pre determined site was an alternative to make I-SceI more programmable. However, a substantial challenge arose since the DNA binding domain could not be decoupled from cleavage domain in naturally occurring I-SceI (Moure et al., 2008).

2.3.2 Zinc-finger nucleases (ZFNs)

ZFNs are artificial proteins that combine the DNA binding domain of a zinc finger protein with the non-specific cleavage domain of a FokI endonuclease, providing a general strategy to deliver site-specific DSBs to the chromosome (Li et al., 1992, Pavletich and Pabo, 1991). A zinc finger (ZF) is a ~30 amino acid motif that recognizes 3 nucleotides of DNA. Three ZF repeats are typically contained in an individual ZFN (Pavletich and Pabo, 1991). To strengthen the interaction of two FokI cleavage domains, a pair of zinc fingers is designed to bind neighboring sequences with a 5 to 7 bp spacer to form a dimeric cleavage domain. Such optimal configuration allows dimerization and subsequent cleavage (Bibikova et al., 2001). Successful ZFN-induced genome modifications were reported in many organisms such as plants (Townsend et al., 2009), zebrafish (Meng et al., 2008), frogs (Young et al., 2011), mice (Carbery et al., 2010), and sea urchin (Ochiai et al., 2010). Potential off-target effects were alleviated by utilizing a less toxic nickase (Kim et al., 2012) or additional ZF repeats (four to six) to bind longer sites (Wood et al., 2011). However, it has been shown that each triplet recognized by a ZF cannot be simply assembled to recognize a longer sequence (Ramirez et al., 2008). As a result, synthesizing customized ZFs remains difficult and expensive (Carroll, 2011).

2.3.3 Transcription activator-like effector nucleases (TALENs)

Similar to ZFNs, TALENs fuse the DNA-binding domain of a transcription activator-like effector (TALE) with a catalytic nuclease domain. TALE is a bacterial effector protein for synergistic regulation of gene expression in Xanthamonas sp. and Ralstonia sp. (Fu et al., 2013). Highly conserved 33–35 amino acid TALE repeat domains each bind one nucleotide of DNA with specificity dictated by two hypervariable residues (Fu, Foden, 2013). Unlike ZFNs, this one-to-one code allows the design of proteins with desired DNA-binding specificities by simply concatenating TALE repeats. Due to the more flexible recognition rule, TALENs can in principle be designed to readily target any sequence across the genome. In fact, TALENs have been applied in various organisms such as human cells (Sun et al., 2012c), yeast (Li et al., 2011) and zebrafish (Li, Huang, 2011). According to some preliminary studies, TALENs also seem to have fewer off-target effects compared to the corresponding ZFNs (Mussolino et al., 2011). Nonetheless, the primary drawback of TALENs is the trickiness of assembling a large number of repeats into an array. Several strategies have been reported to address this limitation (Briggs et al., 2012, Liang et al., 2013).

2.3.4 CRISPR nucleases

The type II bacterial Clustered Regularly Interspaced Short Palindromic Repeats and CRISPR-associated proteins (CRISPR-Cas) system has recently been exploited as an efficient genetargeting technology (Cong et al., 2013, Hwang et al., 2013, Jiang et al., 2013, Wang et al., 2013a). Directed by a trans-activating crRNA (tracrRNA):crRNA duplex (or a chimeric guide RNA (gRNA)), the CRISPR nuclease (e.g., Cas9 protein) is able to cleave a target DNA sequence with the protospacer adjacent motifs (PAM) (Mali et al., 2013). The 12 bp of rigorous homology at 5’ of the PAM sequence ensured the activity of the CRISPR nuclease (Cong, Ran, 2013). Diversified short PAM sequences (e.g., NGG (Deltcheva et al., 2011), NAAR (van der Ploeg, 2009), NGGNG (Horvath et al., 2008)) recognized by different CRISPR nucleases permit almost all sequences to be targeted. Rather than protein based recognition by ZFNs or TALENs, the nucleic acid-based recognition by CRISPR nucleases significantly eases the assembly process. CRISPR-Cas assisted genome modifications have been demonstrated in many species such as mammalian cells (Cong, Ran, 2013, Mali, Yang, 2013), yeast (Dicarlo et al., 2013b), E. coli (Jiang et al., 2013) and plants (Feng et al., 2013). In some cases, the HR and NHEJ efficiencies mediated by the CRISPR nuclease were much higher than those obtained by TALENs (Mali, Yang, 2013). Morever, efficient multiple deletions were also achieved in mammalian cells and S. cerevisiae using CRISPR-Cas (Bao et al., 2014, Wang, Yang, 2013a). However, because of the short recognition sequence (12 bp), many off-target cleavages may be generated by CRISPR nucleases, which may be a serious issue in hosts with large genome sizes. In fact, the off-target problem has been mitigated by several approaches in recent studies. With a pair of closely spaced Cas9 variants (Cas9-D10A) nicking adjacent regions on opposite DNA strands rather than introducing DSBs, the mutation frequencies of off-target sites were reduced in mammalian cells (Cho et al., 2014, Shen et al., 2014). To circumvent the difficulties of adopting the paired nicking strategy for multiplex or genome-scale targeting by CRISPR-Cas, Fu et al. successfully improved the targeting specificity using truncated gRNAs with 17 or 18 nucleotides. Most importantly, those truncated gRNAs can function as efficiently as (or even more efficiently than) their matched 20 nucleotides counterparts (Fu et al., 2014).

2.4 Group II introns

A group II intron (also known as a targetron) consists of a self-catalytic RNA and a ribonucleoprotein (RNP), which catalyzes the insertion of an intron into specific DNA sites via retrohoming (Michel and Ferat, 1995) (Fig. 1D). Through base-pairing with the intron RNA, site-specific insertions or disruptions can be accomplished by group II introns. Group II intron based gene disruptions are prevalent in prokaryotes, albeit there are a few applications in eukaryotes (Jeong, Cho, 2013). Group II introns have been used in bacterial hosts with fairly low HR efficiency (e.g., Clostridia) (Shao et al., 2007). In addition, group II introns can promote targeted insertion, deletion, inversion and cassette exchange by delivery of recombinase recognition sites (e.g., loxP sites) to a given locus (Enyeart et al., 2013). Group II introns were also reported to introduce site-specific DSBs (Karberg et al., 2001), which may facilitate HR or NHEJ-based DNA repair. As observed with other programmable nucleases (e.g., TALENs (Aouida et al., 2014) and CRISPR (Bao, Xiao, 2014)), the targeting efficiency mediated by group II introns was also site-dependent (Perutka et al., 2004).

3 Transcriptome engineering

In addition to genome editing, transcriptome engineering provides a complementary strategy for genome-scale engineering. Targeting at trans-acting regulatory elements, genetic modulation is achieved without modifying the target chromosomal loci. This feature eliminates the need for prior knowledge of host genomes, which is required by most genome editing methods that depend on homologous recombination. Transcription factors (TFs) and regulatory non-coding RNAs (ncRNAs) are the most common targets for transcriptome engineering.

3.1 Transcription factor engineering

Thanks to transcriptional regulatory networks, cells can rapidly coordinate the expression of thousands of genes when facing both internal and environmental stimuli (Lopez-Maury et al., 2008). Such networks exhibit pyramid-shaped hierarchical structures, with most transcription factors (TFs) at the bottom and middle levels, and only a few master TFs on the top for global regulation (Yu and Gerstein, 2006). Whereas specific TFs at the bottom levels modulate dozens of genes in the same functional group, the master TFs have global influence over the gene expression profile (Yu and Gerstein, 2006). These features make TFs the ideal targets for transcriptome reprogramming by modulating many genes simultaneously (Lin et al., 2013, Santos and Stephanopoulos, 2008). Two main strategies have been applied to engineer TFs: modulation of native transcriptional machinery and introduction of artificial TFs (Fig. 1E).

As a demonstration of native TF engineering, global transcriptional machinery engineering (gTME) introduces mutations to the master TFs that mainly mediate DNA recognition, based on the assumption that variations in these TFs may exert substantial changes to the promoter preference of the RNA polymerase. As proof of concept, the principal sigma factor in E. coli (σ⁷⁰) was subjected to error-prone PCR. From the resultant strain libraries, mutants with improved tolerance to sodium dodecyl sulfate (SDS) and ethanol were identified through serial subculturing (Alper and Stephanopoulos, 2007). In S. cerevisiae, the TATA-binding protein Spt15p and a TATA-binding protein-associated factor Taf25p were mutated. The best variant, which harbored three amino acid mutations in Spt15p, conferred a 70% improvement in ethanol productivity (Alper et al., 2006). It has also been demonstrated that gTME is more effective in diversity creation than chemical mutagenesis methods, and therefore increases the possibility to isolate phenotypes that were unattainable through traditional methods (Klein-Marcuschamer and Stephanopoulos, 2008).

On the other hand, artificial transcription factor (ATF) libraries have also been created to generate transcriptional diversities. A minimal ATF may only contain a DNA-binding domain, whose interaction with its target sequence most likely down-regulates the expression of a nearby gene by interfering with transcriptional initiation or elongation (Park et al., 2005). The DNA-binding domain can also be attached to effector (activator/repressor) domains or ligand-binding domains, which permits more sophisticated regulation. Most ATFs reported so far have employed zinc finger proteins (ZFPs) as the DNA-binding domains, and the library of ATFs is constructed through combinatorial assembly of individual zinc fingers with diverse DNA-binding specificities. The first example of such effort was the introduction of over 10⁵ ZFPs fused with effector domains into S. cerevisiae (Park et al., 2003). The library consisted of threeor four-finger ZFPs, which recognize 9 bp or 12 bp DNA sequences with limited randomness constrained by choice of individual zinc fingers (40 and 25 individual zinc fingers for three-and four-finger proteins, respectively). Several ATFs were identified to confer a number of tolerance phenotypes towards heat, osmotic pressure and an antifungal drug ketoconazole. The relatively short recognition sequence (9 bp or 12 bp) permits the ATFs to modulate many genes. For example, one selected artificial transcriptional factor (K7), which conferred ketoconazole resistance, had 14 perfectly matched binding sites in the yeast genome (Park, Lee, 2003). The perturbation scope by these AFTs may be even larger considering off-target effects. A similar strategy has been applied to E. coli to isolate tolerant strains towards heat shock (Park, Jang, 2005) and butanol (Lee et al., 2011). For future development, we envision that a new generation of ATFs can be developed when the DNA-binding domains are changed from ZFPs to TALEs and CRISPR proteins. As the DNA-binding specificity of TALEs and CRISPR proteins is much more predictable than that of ZFPs, transcriptome perturbation by TALE- and CRISPR-derived ATF libraries should be more effective and programmable.

3.2 Regulatory non-coding RNAs

Non-coding RNA molecules (ncRNAs) are increasingly recognized as key regulators across the biological kingdoms (Kang et al., 2014, Qi and Arkin, 2014). Here we will mainly focus on regulatory ncRNAs that have been used in genome-scale engineering (Fig. 1F). To be suitable for genome-wide applications, synthetic ncRNAs should be preferably trans-acting, permitting simple introduction of a genome-wide library with minimal considerations on local genetic context. Also, the interaction between an ncRNA and its DNA or mRNA target should be mainly determined by Watson-Crick base pairing, so that the binding specificity and efficiency can be predictable and programmable.

In bacteria, trans-acting small RNAs (sRNAs) and antisense RNAs (asRNAs) are two main regulatory ncRNAs (Qi and Arkin, 2014). Lee and coworkers recently developed a general framework to design synthetic sRNAs in E. coli for metabolic engineering (Na et al., 2013). The synthetic sRNAs were composed of a scaffold sequence and a target-binding sequence. The scaffold was derived from a naturally occurring sRNA, MicC, and the scaffold can recruit the Hfq protein to facilitate sRNA-mRNA interaction and mRNA degradation. The native target binding sequence of MicC can be replaced by the antisense sequence to the translation initiation region (TIR) of any given gene. Correlation was found between the repression capability and the binding energy of the antisense sequence, which allowed for fine-tuning of the knockdown efficiency. Although the sRNA library constructed in that study only targeted the cadaverine production related genes (Na, Yoo, 2013), it is possible to expand the strategy to a genome-scale.

On the other hand, asRNAs have been used for functional genomics study in a series of bacteria, such as Streptococcus mutans (Wang and Kuramitsu, 2005) and Staphylococcus aureus (Forsyth et al., 2002). However, it has been long recognized that asRNAs are inefficient for gene repression in E. coli (Wagner and Flardh, 2002). Recently, it was found that asRNA molecules with paired-termini have enhanced stability and improved repression capacity (Nakashima et al., 2006). E. coli genomic DNA fragments have been cloned into a paired-termini expression vector to generate a genome-wide asRNA library, which was used to successfully isolate asRNAs that target essential genes and led to conditional growth inhibition (Meng et al., 2012).

As for eukaryotes, the most common ncRNA machinery for gene expression regulation is RNA interference (RNAi), a cellular gene silencing mechanism whereby mRNAs are targeted for degradation by homologous double-stranded RNAs (dsRNAs) (Fire et al., 1998, Hannon, 2002). RNAi proves to be a powerful tool for genome-wide reduction-of-function screen in many higher eukaryotes (Boutros and Ahringer, 2008, Echeverri and Perrimon, 2006), yet its applications in microbes are rare. This is probably due to the lack of a native RNAi pathway in S. cerevisiae, which is the most-widely used microbial eukaryote. Recently, a heterologous RNAi machinery has been reconstituted in S. cerevisiae (Drinnenberg et al., 2009), which opens up the possibility of genome-wide RNAi screen (see Section 6.3).

In addition to naturally-occurring ncRNAs, synthetic RNAs were also used to modulate gene expression in the CRISPR-mediated interference (CRISPRi) and CRISPR-mediated activation technologies (Fig. 1F). By co-expressing of a Cas9 mutant with abolished endonuclease activity (dCas9) and a guide RNA targeting at the non-template DNA strand of a target gene, up to 1,000-fold reduction in gene expression was achieved in E. coli (Qi et al., 2013). In S. cerevisiae, the silencing efficiency may be further improved by fusing the dCas9 protein to transcription repressors or chromatin silencers (Gilbert et al., 2013). For gene activation, transcriptional activators can be delivered by the dCas9-guide RNA complex to the upstream region of a promoter, resulting in up-regulation in E. coli (Bikard et al., 2013) and yeast (Farzadfard et al., 2013, Gilbert, Larson, 2013).

4. Genome synthesis

Genome synthesis is one of the most impressive achievements of synthetic biology, ranging from viral genomes (Cello et al., 2002, Chan et al., 2005) and bacterial genomes (Gibson et al., 2008, Gibson et al., 2010, Karas et al., 2012) to yeast chromosomes (Annaluru et al., 2014, Dymond et al., 2011). Early efforts mainly focused on increasing the scale of the final DNA constructs, from the 7.5 kb cDNA copy of a poliovirus genome (Cello, Paul, 2002) to a 1.08 Mb bacterial genome (Gibson, Glass, 2010). In terms of DNA sequences, the synthetic genomes are almost exact copies of the native ones, except for a few inserted “watermarks” such as the names of the team members (Gibson, Benders, 2008, Gibson, Glass, 2010). A recent report took a step further to build a designer yeast chromosome that was substantially different from its wild-type template (Annaluru, Muller, 2014). Compared to the native chromosome III of S. cerevisiae, the designer chromosome synIII was ~ 14% smaller due to the deletion of some non-essential regions such as transfer RNAs, transposons and introns. However, the design principles are still very simplistic, and it requires substantial technological development before we can write a fully synthetic genome. Therefore, rather than discussing the practice of genome synthesis, we will instead focus on its enabling technologies, such as DNA synthesis and parts engineering, as these technologies are also highly useful in engineering better microbial cell factories.

4.1. DNA synthesis

One fundamental enabling factor for synthetic genomes is the decreasing cost of chemical DNA synthesis in a manner akin to Moore’s law (Mueller et al., 2009). While further reduction of the synthesis cost using the traditional column-based methods is unlikely, microarray-based technologies provide great potential for high-throughput and cost-effective DNA synthesis, as discussed in details elsewhere (Mueller, Coleman, 2009, Tang et al., 2013). Here we will focus on two recent advances in microarray-based DNA synthesis, highlighting the importance of technology integration. One major limitation of chip-synthesized sequences is that they are prone to error. In addition to the optimization of the microarray technology itself, the integration of next-generation sequencing (NGS) has greatly improved the fidelity of synthetic DNA (Matzas et al., 2010). A pool of chip-derived oligonucleotides were attached to individual streptavidin-coated beads and amplified via emulsion PCR. Pyrosequencing was performed, and the beads with correct sequences were sorted in a high-throughput manner by a camera-guided micropipette. The fidelity was estimated to be improved by 500-fold, and the throughput may potentially enable DNA construction at a megabase scale (Matzas, Stahler, 2010). Another challenge lies at the heterogeneity of the microarray-generated oligo pools, as well as the limited amount of DNA synthesized at an individual spot, both of which complicate the subsequent gene assembly. Quan et al. devised an integrated solution to overcome this challenge, by combining the synthesis, amplification and assembly steps (Quan et al., 2011). The microchip was divided into microfluidics aided subarray reactors, each containing oligonucleotides for the assembly of the same gene. After synthesis, the oligos were amplified, released and assembled into gene constructs up to 1 kb each by enzymatic reactions. Together, both methods discussed above demonstrate that novel solutions can be achieved through technology integration, which will continue to be a driving force to provide synthetic DNA with lower price and higher quality.

4.2. Parts engineering

With the DNA synthesis capacity becoming less restrictive, the limited repertoire of genetic parts and a lack of design principles are becoming the two major obstacles in synthetic genome construction (Wang et al., 2013b). Rather than providing comprehensive summary on how to overcome these obstacles, we will only focus on the expansion of part collections, which also provides sources for diversity generation to improve MCF performance. The readers are directed to other excellent reviews for the design framework for functional assembly of biological parts (Wang, Wei, 2013b, Way et al., 2014), as well as experimental protocols for physical assembly of DNA fragments (Chao et al., 2014).

Based on their functional roles, genetic parts can be classified as sensors (take in environmental stimuli, e.g. riboswitches), regulators (perform calculation, e.g. promoters), actuators (generate outputs, e.g. structural proteins) and adapters (connect components, e.g. endoplasmic reticulum (ER) tags) (Wang, Wei, 2013b). Here we will mainly discuss transcriptional regulators, such as promoters, RBSs and terminators, as they are so far still the most available, well-studied and widely-used genetic parts. Given the wide application of microbial biotechnology, the most important task on these building blocks is to discover, engineer and characterize as many biological parts as possible (Wang, Wei, 2013b). There are mainly three methods to enlarge the parts collection: to harvest from nature, to create mutant libraries, and to build by modeling.

Advances in sequencing and bioinformatics permit rapid identification and prototyping of biological parts from genome and megagenome sequences. For example, strong constitutive promoter candidates can be isolated from the upstream sequences of housekeeping genes (such as global transcription/translation factors, glycolytic enzymes, etc) (Shao et al., 2013, Sun et al., 2012a). High-throughput techniques may also allow characterization of all putative promoters in a microbe (Zaslaver et al., 2006). To create a mutant library of biological parts, there are essentially three strategies. First, mutations can be introduced via error-prone PCR, and the resultant variants with desirable performance (e.g. various promoter strengths) are isolated through high-throughput methods (Alper et al., 2005, Nevoigt et al., 2006). Second, the heterogeneous DNA oligonucleotide pools synthesized by microarray can serve as a source for diversity. For example, synthetic intergenic regions were combinatorially assembled from three oligo libraries with overlapping ends, modulating relative expression patterns of multiple enzymes in a synthetic operon (Pfleger et al., 2006). Also, recombineering enabled the replacement of the native ribosome binding sites (RBSs) with degenerate libraries in E. coli (Wang, Isaacs, 2009). Finally, chimeric parts can be constructed by inserting modulating cassettes into native parts. For example, arrays of upstream activation sequences (UASs) were placed in front of a core promoter element to create strong synthetic promoters in S. cerevisiae (Blazeck et al., 2012) and Yarrowia lipolytica (Blazeck et al., 2011). If the genetic basis that controls the part performance is relatively clear, modeling can help guide the construction of synthetic parts predicted parameters (Crook and Alper, 2013). For example, a computational model has been established to predict the translational initiation rate of bacterial RBSs, by quantifying the interactions between the 30S ribosome complex and the target mRNA molecule (Salis et al., 2009). Based on a nucleosome architecture model, purely synthetic yeast promoters were obtained with decent strengths (Curran et al., 2014).

5 High-throughput genotype-phenotype mapping

With limited understanding of complex biological systems, “rational” genetic engineering often encounters challenges. Instead, “inverse metabolic engineering” (IME), which isolates mutant strains with a desirable trait first and then proceeds to determine underlying genetic changes, has proved to be a more effective strategy (Bailey et al., 2002, Santos and Stephanopoulos, 2008). IME not only rapidly improves a target phenotype, but also provides insights to guide future engineering efforts. To fully realize the potential of IME, advanced genotyping and phenotyping techniques are needed to expedite the cycles of creating diversity, selecting best mutants, and mapping relevant genetic changes (Garst et al., 2013). Genotyping helps to create and track comprehensive and/or combinatorial diversity across the genome, whereas phenotyping helps to identify mutants with desirable traits in a high-throughput manner.

5.1. Genotyping

5.1.1. Genome-wide libraries

Genome-wide overexpression/knockout libraries are powerful tools to comprehensively investigate the impact of individual genetic modification on a given phenotype (Fig. 2). For overexpression libraries, either genomic fragments (Lynch et al., 2007) or all the open reading frames (ORFs) under the control of a promoter (Ho et al., 2009) can be cloned into an extrachromosomal vector. After screen/enrichment, the inserts can be identified through microarray analysis (Lynch et al., 2004) or DNA sequencing. Genome-wide knockout libraries can be generated by many strategies. Transposon mutagenesis has been optimized for unbiased integration of an antibiotic marker cassette into the entire genome, hence creating a random knockout library (Alexeyev and Shokolenko, 1995, Badarinarayana et al., 2001). Moreover, all the nonessential genes can be disrupted through homologous recombination, examplified in construction of the yeast deletion collection (Giaever et al., 2002, Winzeler et al., 1999) and the Keio E. coli knockout collection (Baba et al., 2006). In addition to knockout, reduction-of-function screen has also been applied for genome-wide analysis. Examples include the above-mentioned screening with as-RNAs and RNAi. Knockdown libraries are especially important to study essential genes whose deletion mutations are lethal. For example, insertion of an antibiotic marker into the terminator region to destabilize the mRNA enabled knockdown modification on the essential genes in S. cerevisiae (Breslow et al., 2008). Notably, there are also technologies that can create comprehensive genetic libraries including both overexpression and knockdown modifications. For the trackable multiplex recombineering (TRMR), two kinds of synthetic cassettes were designed for promoter replacement: the ‘up’ cassette containing a strong promoter, and the ‘down’ cassette containing an inert sequence to replace the native RBS. Through recombineering, these synthetic cassettes were incorporated in front of every gene in E. coli, which led to either increased or decreased expression of a target gene (Warner, Reeder, 2010). Though not demonstrated to create microbial genome-wide libraries yet, CRISPR-mediated knockout, interference and activation can be readily applied for genome-scale analysis as discussed previously.

Fig. 2. Genome-wide strain libraries for high-throughput genotyping.

Single mutations can be introduced through plasmid-borne libraries or directed genome editing. A double-mutation strain library can be created using the synthetic genetic array (SGA) method, whereby a query strain harboring the first mutation can be mated with a strain library to incorporate a genome-wide second mutation.

Adding more dimensions to such approaches, i.e. modification at two or even more loci, is necessary because of the non-linear interactions between single genetic variations (Fig. 2). For combinatorial overexpression libraries, the coexpressing genomic libraries (CoGEL) approach was used to construct genomic libraries in a series of vectors (plasmid or fosmid) with compatible replication origins and different resistance markers, which enabled coexistence of two or more genomic inserts in one cell. This approach successfully identified known and novel combinations of genetic changes that conferred improved acid tolerance in E. coli (Nicolaou et al., 2011). On the other hand, construction of a double-mutant library from single loss-of-function collections by mating or conjugation has been demonstrated in model organisms such as E. coli (Butland et al., 2008, Typas et al., 2008) and S. cerevisiae (Pan et al., 2004, Tong et al., 2001). An impressive application was the depiction of a genome-scale digenic interaction network in S. cerevisiae, by examining 5.4 million gene-gene pairs in a double-mutant library (Costanzo et al., 2010). However, current protocols to generate genome-wide double-mutant libraries are quite resource-intensive and time-consuming, as complicated replica-pinning procedures are needed to perform mating, recombination and selection. Therefore, alternative approaches that simplify the introduction of a second mutation on a genome-scale are desirable to speed up the discovery of synergistic modifications. For example, the inherent multiplex capacity of the CRISPR system can be used to create combinatorial genome-wide libraries. As a first step towards this objective, we recently developed a homology integrated CRISPR-Cas (HI-CRISPR) system for one-step multigene disruptions in S. cerevisiae. The mutagenizing homologous recombination donor is integrated at the 5’ of the guide sequence. To increase the gene disruption efficiency, all the HI-CRISPR elements are embedded on a plasmid with ultrahigh copy number. As proof of concept, the simultaneous disruption efficiency of three genes (CAN1, ADE2 and LYP1) ranged from 27% to 87%, which enabled the identification of desired mutants by random genotyping (Bao, Xiao, 2014).

To facilitate subsequent analysis with these genetic libraries, molecular barcodes have been used to monitor the abundance of every mutant strain in a mixed population. Microarray analysis with complementary probes (Pierce et al., 2006), as well as the “Bar-Seq” method with NGS (Smith et al., 2009), can be used to quantify the dynamics of barcodes and their linked mutants in various screening experiments, enabling high-throughput mapping of relevant genes to a given phenotype. Such high-throughput capacity explains the wide application of molecular barcodes in analyzing overexpression (Ho, Magtanong, 2009), knockout (Giaever, Chu, 2002) and TRMR libraries (Warner, Reeder, 2010).

5.1.2. Genome and transcriptome analysis by next-generation sequencing

Next-generation sequencing (NGS) offers a rapid and inexpensive way to perform genome-scale analysis, whose recent technological advances are reviewed elsewhere (Mardis, 2008, 2013). Here we emphasize on two techniques, comparative genomics (Borneman et al., 2013) and comparative transcriptomic profiling by RNA-seq (Mutz et al., 2013, Ozsolak et al., 2009), which are highly informative in revealing genotype-phenotype relationship, genetic interactions and regulatory networks in microbes.

Comparative studies of genome sequences among laboratory strains, industrial strains and natural isolates of the same species can provide insights on the genetic basis of phenotypic differences (Borneman, Pretorius, 2013). For example, the genomes of four wine and two brewing strains of S. cerevisiae have been sequenced and analyzed against existing S. cerevisiae genome sequences. The variations between different S. cerevisiae strains include single-nucleotide polymorphisms (SNPs), insertions and deletions, and as novel, strain and allele-specific ORFs, and genomic rearrangements (Borneman et al., 2011). Transcriptomic profiling is useful to reveal the impact of different environmental factors on gene expression. For example, transcriptomes of Bacillus subtilis under 104 different environmental and nutritional conditions has been investigated to reveal the structure of transcription network, by grouping 2935 promoters into regulons and then linking these regulons with various transcription factors (Nicolas et al., 2012). For a parental strain and its derivatives obtained from evolutionary engineering or genetic engineering, both genomic and transcriptomic analysis are helpful to understand the genetic basis of acquired traits. Whole genome sequencing (WGS) has been used to track the evolutionary trajectory for aerobic citrate utilization in E. coli (Blount et al., 2012) and efficient xylose fermentation in Scheffersomyces stipitis (Smith et al., 2008). Specifically, mutations that do not affect gene expression can only be identified by genome sequencing. For example, an S. cerevisiae mutant strain with the only lactate transporter gene JEN1 deleted regained the ability to grow on lactate as the sole carbon source during adaptive evolution. Transcriptome analysis provided no clues on the evolved strain; the single-nucleotide changes that conferred an acetate transporter Ady2p with lactate transport activity were only discovered by genome sequencing (de Kok et al., 2012). On the other hand, transcriptional profiling is important to understand the mechanisms on how genetic changes affect a given trait, especially for those mutations resulting in large-scale perturbation in gene expression. For example, differential expression of hundreds of genes caused by mutations in TFs can only be revealed by transcriptome analysis (Alper, Moxley, 2006).

5.2 Reporter-based phenotyping

In addition to generating and detecting modifications on a genome-scale, rapid identification of mutants with desired properties from genome-wide libraries is required in IME. In fact, many high-throughput phenotyping methods, including microplate screening (Behrendorff et al., 2013), surface display (Boder and Wittrup, 1997), compartmentalization (Wang et al., 2014) and fluorescence-activated cell sorting (FACS) (Santoro and Schultz, 2002), have been exploited to detect phenotypes that can be directly screened or selected for (e.g., fluorescence and cell survival, respectively). These methods have been reviewed elsewhere (Leemhuis et al., 2009) and will not be discussed here. However, phenotyping properties that are not amenable to high-throughput screening and/or selection (e.g., bulk and fine chemicals production), remains challenging. To address this limitation, a variety of reporter-based phenotyping methods have been developed to link easily detectable phenotypes to the desired traits.

In vivo production of a target molecule can be monitored via a TF-promoter based reporter system. Under the regulation of the molecule-responsive TF and its cognate promoter, expression of the reporter gene results in colorimetric, fluorescent or growth-coupled phenotypes, which are correlated to the concentration of the target molecule and are rapidly identifiable (Fig. 3). Using TF-promoter pairs from different microbes, specific activation of transcription factors by dicarboxylic acids and alcohols was coupled to expression of the tetracycline reporter gene. As proof of concept, a biosensor was constructed to identify E. coli mutants with 35% higher specific productivity of 1-butanol. This biosensor was also incorporated in a synthetic selection method that couples 1-butanol biosynthesis with cell fitness, leading to a 120-fold increase in 1-butanol production (Dietrich et al., 2013). On the other hand, a feedback-regulated evolution of phenotype (FREP) system was developed to control mutation rates, which are thought to be associated with the probability of finding improved mutants and therefore speed up the adaptive evolution process (Metzgar and Wills, 2000). In the absence of the target molecule, the sensor assembled from the TF-promoter pair activates expression of the actuator and reporter. Expression of the actuator results in an increase in the mutation rate, which may lead to increased production of the target molecule. In response to the increased concentration of the target molecule, the sensor cannot activate the expression of the actuator and reporter, giving rise to a decrease in mutation rate and thus an increase in hereditary stability. Several synthetic E. coli TF-promoter pairs were accordingly engineered for better response to the metabolic intermediate isopentenyl pyrophosphate (IPP). Adopting such sensors in FREP was able to increase tyrosine and isoprenoid production in E. coli (Chou and Keasling, 2013). With a wide dynamic range, high sensitivity and specificity towards a given ligand, the well-characterized ligand-responsive TF and promoter pair always tends to be adopted in an efficient reporter based screening. Yet, promoter characterization is scant in prokaryotes and almost vacant in eukaryotes to date. Also, the most commonly used TF and promoter pairs are unlikely to be responsive to majority of industrially relevant molecules (Dietrich et al., 2010).

Fig. 3. Reporter based phenotyping.

In the presence of a ligand, the specific activation of transcription factor (TF) is controlled by the expression of the reporter gene. A synthetic riboswitch acts as a biosensor for desired metabolites.

Due to the lack of effective TF-promoter pairs, synthetic riboswitches responsive to various industrially relevant compounds have become an alternative choice for reporter based screening in eukaryotic organisms. Synthetic riboswitch contains an input domain (encoded in an RNA aptamer) and an output domain (encoded in a ribozyme), which is placed in the 3’ untranslated region of a reporter gene. In the absence of a ligand, self-cleavage is catalyzed by the output domain, leading to low gene expression. Binding to a desired molecule by the input domain results in misfolding of the output domain, leading to low cleavage activity and thus high gene expression (Fig. 3). For example, coupled with FACS, an engineered riboswitch was used to identify caffeine demethylase mutants with significantly improved activity and product selectivity (Michener and Smolke, 2012).

6 Notable examples

Recent advances in genome-scale engineering have significantly enhanced the ability to generate and map multiple functional changes across the entire genome. As such, impressive progress in rewiring genomes to elicit robust, complex traits has been achieved with minimal prior knowledge of the genetic determinants. A few notable examples of genome-scale engineering will be highlighted in this section.

6.1 MAGE

Based on the λ Red recombination system, MAGE uses multiple oligos to create combinatorial libraries and optimize gene expression (Fig. 4). Most importantly, the recursive cycles of oligo introduction and allelic replacement are carried out by an automated system, enabling generation of over 4.3 billion genetic variants per day. In one study, the RBSs of 24 genes related to the 1-deoxy-D-xylulose-5-phosphate (DXP) pathway were concurrently modified to increase lycopene production. Mutants with over 5-fold increase in lycopene production were isolated within 3 days (after 35 MAGE cycles), representing a significant improvement over previously reported efforts (Wang, Isaacs, 2009). In another study, MAGE was exploited to insert short DNA sequences into the E. Coli choromosome. After 110 MAGE cycles, 18-nt hexa-histidine tag sequences were successfully inserted into 38 essential genes encoding the entire translation machinery, allowing modification and co-purification of large protein complexes and pathways (Wang et al., 2012).

Fig. 4. Multiplex automated genome engineering (MAGE) and trackable multiplex recombineering (TRMR) accelerated E. coli genome evolution.

MAGE enables rapid generation of sequence diversity via continuous delivery of ss-oligos into cells. With barcode incorporated oligos, TRMR enables simultaneous creation and tracking of multiple genetic modifications.

Meanwhile, various MAGE-derived methods were further developed for efficient genome editing. For example, the use of “coselection” MAGE (CoS-MAGE) greatly improves the recombineering efficiency. In the MAGE-generated combinatorial variants, a small portion of cells were observed to harbor multiple mutations (Isaacs, Carr, 2011), which could be selected out in the presence of selective markers. By leveraging co-selection markers around the target site, oligo-mediated allelic replacement efficiency of over 70% per viable progeny was achieved (Carr et al., 2012). Consequently, this approach was used to insert T7 promoter sequences to 12 genomic operons related to aromatic amino acid biosynthesis, permitting rapid generation of promoter libraries (Carr, Wang, 2012). A recently-reported microarray oligonucleotide-MAGE (MO-MAGE) method can cost-effectively amplify thousands of oligos from microarray chips. By adopting such technology, T7 promoters were inserted to the upstream of 2585 operons with an average frequency of 0.02% per locus and 0.4 average insertions per cell (Bonde et al., 2014). Additionally, the hierarchical conjugative assembly genome engineering (CAGE) method was investigated to reprogram the genetic code. The 32 regions of codon modifications (replacing TAG stop codons with TAA) constructed by MAGE were merged into a single genome through conjugation. Because the TAG stop codon was absent in the resultant strain, this liberated codon can be reassigned to a novel amino acid (Isaacs, Carr, 2011).

6.2 TRMR

To reduce the unwanted secondary mutations produced by excessive cycles, MAGE is often limited to create modifications to a subset of relevant genes, which requires a prior knowledge of which genes are to be targeted. A complementary method, TRMR, provides a clue to address this limitation. After the introduction of oligos with unique barcodes for recombineering, cells with desired mutations were enriched in a favorable environment and the corresponding genetic modifications could be quantitatively tracked using the barcoded sequences and microarray analysis (Warner, Reeder, 2010) (Fig. 4). As such, thousands of genes that affected E. coli growth in rich, minimal or cellulosic hydrolysate media in the presence of β-glucoside, D-fucose, valine and methylglyoxal were mapped within one week, permitting identification of large sets of targets for genome engineering endeavors (Warner, Reeder, 2010). TRMR can also assist MAGE to achieve directed genome engineering via identifying the most relevant genetic modifications. By coupling of TRMR with MAGE, barcoded promoter mutant libraries were first introduced to modify gene expression under different challenging environments. Based on the mapping results, RBS mutant libraries were then designed to retarget genes that dramatically affected the cell growth. Consequently, extensive growth-enhancing mutations were identified from different conditions (Sandoval et al., 2012).

6.3. RAGE

Due to the lack of an efficient recombineering mechanism, it is difficult to apply MAGE or TRMR for genome-scale engineering in eukaryotic microorganisms such as S. cerevisiae (DiCarlo et al., 2013a). On the other hand, RNAi screening has been widely used for functional genomics research in various eukaryotes (Boutros and Ahringer, 2008, Echeverri and Perrimon, 2006), yet its applications in MCF engineering are still rare. Recently, a heterologous RNAi pathway was reconstituted in S. cerevisiae (Drinnenberg, Weinberg, 2009), which enables the use of RNAi screening to rapidly understand and engineer complex phenotypes in this yeast (Fig. 5). By inserting random genomic DNA fragments into a pair of convergent constitutive promoters, double-stranded RNAs were transcribed in vivo to elicit genome-wide knockdown in the presence of the RNAi pathway (Si et al., 2014). The resultant library was used to successfully identify known suppressors of a telomere-defect mutation yku70Δ and genetic determinants for improved resistance towards acetic acid and furfural (Si, Luo, 2014, Xiao and Zhao, 2014). Moreover, compared with the traditional conjugation-based method (Tong, Evangelista, 2001), plasmid-borne RNAi screening is much more convenient in creating genome-wide perturbations in a modified strain background. Therefore, it is possible to apply a directed evolution strategy on a genome-scale to engineer complex phenotypes in S. cerevisiae (Fig. 5). For example, RNAi-assisted genome evolution (RAGE) was developed to identify three knockdown mutations that acted synergistically to improve acetic acid tolerance substantially in S. cerevisiae (Si, Luo, 2014).

Fig. 5. RNAi-assisted genome evolution (RAGE) in S. cerevisiae.

In the presence of a heterologous RNAi pathway, genome-wide knockdown screening can be performed with a double-stranded RNA library derived from genomic DNA. Iterative RNAi screen may help to accumulate beneficial genetic modifications in an evolving yeast genome for continuous improvement of a complex phenotype.

7 Conclusions and perspectives

Recent advances in genome-scale engineering have overcome many barriers that constrain strain engineering and therefore greatly expanded our ability to reprogram biological systems. Owing to their high-throughput, complexity, fidelity, and low cost, these new genome-scale engineering technologies have been able to quickly generate vast libraries of combinatorial variations that may have a greater effect on a given phenotype as well as rapidly map the desired trait. Currently these approaches have heretofore been implemented largely in certain model organisms, but adapting them to other industrial microbes is highly desirable. However, this is not a simple task, as the newly developed methods require effective transformation protocols and certain sets of genetic tools, both of which are often absent for industrial fermentation hosts. Whereas establishing genetic manipulation toolbox in less studied organisms will expand the application of genome-scale engineering in industrial settings, high throughput genotyping and phenotyping technologies, such as whole-genome sequencing, transcriptional profiling, and microfluidic and robotic screening, can be combined with classical strain engineering efforts to decrease the length of time for isolation of improved variants and analysis of underlying mechanisms (Crook and Alper, 2012).

As an alternative strategy, rather than dealing with a large number of variants, genome-scale metabolic models can be used to narrow the search space and create functionally rich libraries for optimizing complex traits. For example, employing designs predicted by in silico models has led to not only enhanced production of desired compounds (Choi et al., 2010, Park et al., 2007), but also discovery of novel drug targets (Kim et al., 2011, Lewis et al., 2010). For a metabolic model, the ability to correctly predict the physiological characteristics of the organism is very important. Hereby, extensive experimental data including high-throughput omics data have been incorporated to validate and improve the quality and the accuracy of a genome-scale metabolic model (Plata et al., 2010). In such a scenario, the massive data obtained from genome-scale engineering could be significant for construction of more refined and complex metabolic models, which may further benefit the genome-scale engineering in return.

Furthermore, laboratory automation may greatly accelerate microbial genome-scale engineering. By eliminating human intervention, laboratory automation promises to improve productivity and reliability, increase throughput, and reduce experimental error rates due to human factors (Linshiz et al., 2013). Pharmaceutical industry has heavily relied on automation technologies to identify new drug lead compounds by screening small-molecule libraries (Nettekoven and Thomas, 2002). Automation platforms have also been developed for bottom-up construction of genetic circuits and metabolic pathways from modular parts (Densmore and Hassoun, 2012, Dharmadi et al., 2014), integrating software for assembly algorithm and data management, as well as hardware for liquid handling and DNA construct analysis. For biological system engineering, the promise of automation has been demonstrated by several recent examples (Esvelt et al., 2011, Wang, Isaacs, 2009). Phage-assisted continuous evolution (PACE) executed 200 rounds of protein evolution in 8 days, during which targeted activities effectively emerged from undetectable levels (Esvelt, Carlson, 2011). Moreover, MAGE created over 4.3 billion combinatorial variants per day, enabling 5-fold increase in lycopene production in 3 days (Wang, Isaacs, 2009). Future development of automation-friendly protocols, including both strain variant generation and phenotypic screening methods, may help to increase the use of laboratory automation in genome-scale engineering of microbial cell factories. Finally, we envision that, with endeavors on but not limited to the above mentioned aspects, genome-scale engineering can accomplish intensive reprogramming of microbial metabolism for multiple engineering purposes.