High-throughput evaluation of synthetic metabolic pathways

Abstract

A central challenge in the field of metabolic engineering is the efficient identification of a metabolic pathway genotype that maximizes specific productivity over a robust range of process conditions. Here we review current methods for optimizing specific productivity of metabolic pathways in living cells. New tools for library generation, computational analysis of pathway sequence-flux space, and high-throughput screening and selection techniques are discussed.

Introduction

Microorganisms have the potential to produce many chemicals of use to society^1,2. In some cases, production from heterologous microorganisms is more sustainable than purifying the chemical from natural sources. Examples include harvesting Pacific yew trees or Chinese wormwood for taxol³ or the anti-malarial artemisinin⁴, respectively. Additionally, the ability to create renewable and sustainable biofuels and biochemicals is increasingly attractive given concerns about climate change and peak oil⁵.

An organism producing a desired product may not exist, or a given strain may not be suitable for required economical processing conditions. Because of this, reconstructed pathways are often implanted into chassis microorganisms⁵. Some of these pathways include those specific for biofuels (ethanol⁵, isobutanol⁶, 1-butanol⁷, 1,4-butanediol⁸), polymer monomers (polylactic acid⁹, isoprene¹⁰, 3-hydroxypropionic acid¹¹) and pharmaceutically active ingredients (precursors for taxol³ or opioids¹²). However, in many cases product toxicity or transport limits end titers, product recovery from aqueous fermentation broths is inefficient, or the volumetric productivity is below that required for a cost-effective process. Combined, these limitations temper the promise of sustainable replacement of the palette of petrochemicals and naturally extracted specialty chemicals currently in use by society.

In particular, the specific productivities of most engineered metabolic pathways are far below that which is needed for industrial production. Some implanted pathways have limited flux because of substantial thermodynamic reversibility at key steps¹³. Additionally, pathway enzymes transplanted into heterologous hosts often have poor expression because of weak catalytic efficiency¹⁴, poor protein solubility, or membrane targeting issues^12,15. Host-specific problems include cofactor accessibility¹⁶, siphoning of pathway intermediates, intermediate toxicity and post-translational flux regulation of key precursors^12,17,18. Furthermore, the performance of a specific engineered metabolic pathway may differ between host strains¹⁹, media formulations¹², temperatures and oxygen conditions²⁰.

A grand challenge in the field of metabolic engineering is the accurate and efficient identification of a pathway genotype that maximizes specific productivity over a robust range of process conditions. Attempts to improve specific productivity have largely focused on screening individual pathway enzymes for activity or balancing gene expression by testing libraries of elements like promoters and ribosome binding sites (RBS)^21–24. Error-prone polymerase chain reaction (PCR) mutagenesis of pathway enzymes has also been used to find activity-improving mutations^17,18. However, pathway optimization by total enumeration becomes unwieldy, as balancing activity at multiple nodes leads to a combinatorial explosion. Consider a plasmid-encoded pathway composed of a series of expression elements (e.g. promoters, RBS, terminators) and pathway gene variants (Fig. 1a). A pathway library comprised of a single enzyme of average length²⁵ driven by 10 alternative promoter and 10 alternative RBS sequences can be covered by testing 10² variants. A library containing the above gene expression genotypes with all possible single non-synonymous mutations to the enzyme now contains 6 × 10⁵ variants. Testing the same number of variants using a two-enzyme pathway requires a theoretical coverage of 3.6 × 10¹¹ variants, which is too large a sequence space to cover under most conditions. The combinatorial problem only becomes more acute with more pathway enzymes. To partially circumvent this combinatorial intractability, modular pathway design has been used to partition individual enzymatic steps into reaction groups. Then, expression of the resulting modules are balanced, not individual enzymes^12,26–28 (Fig. 1b). Alternative ways to explore sequence-flux space include computational predictions from small training sets^29–31 (Fig. 1c) or high-throughput screening or selection techniques^32–36 (Fig. 1d).

Figure 1.

There are three main strategies to sample sequence-flux space in metabolic pathways. (a) A model pathway. Each arrow represents a different reaction chemistry used to convert an initial substrate into a value-added product. Expression elements like promoters and ribosome binding sites (RBS) facilitate expression of the enzyme-encoding gene sequences. (b) Individual reaction chemistries are grouped into modules. Expression of these modules is varied, reducing the combinatorial search space. (c) Given a training set of different sequences with a given output, models can be harnessed to predict optimal pathway expression levels. (d) Advanced DNA assembly methods can be used to create unique pathway variants that are then assessed using high-throughput screens or selections.

The focus of this review is on new technologies that identify highly productive and robust synthetic metabolic pathways. This review will not cover continuous evolution³⁷, whole-genome engineering³⁸ or computational pathway design³⁹ — the interested reader can find excellent reviews on some of these topics elsewhere^40–43. We begin by describing pathway evaluation of isogenic cultures. We next describe computational approaches to predict high-performing genotypes with a limited training set of pathway variants. Next, we consider high-throughput methods to assess metabolic pathways, including population-based screens or selections. New enabling techniques for DNA library construction, sequencing and evaluation will be described throughout.

Evaluation of Isogenic Cultures

One way to evaluate pathway variants is through the use of isogenic cultures. In a typical set-up, a library containing a combinatorial library of expression elements or enzyme variants is created, and clonal variants are tested individually. Lu et al. optimized a xylose fermentation pathway in Saccharomyces cerevisiae by shuffling promoters of various strength in front of each pathway enzyme⁴⁴. Different promoter combinations were made and tested individually for ethanol productivity and enzymatic activity. Solomon et al. tested different expression levels of glucokinase (glk) and galactose permease (galp) to enable glycolytic uptake in Escherichia coli independent of the phosphotransferase system⁴⁵. Carbon flux was modulated by controlling expression of glk and galp under control of synthetic constitutive promoters. Juminaga et al. constructed a pathway for L-tyrosine production in E. coli MG1655 by modifying plasmid copy numbers, promoter strength, gene codon usage, and placement of genes in operons⁴⁶. The best pathway variant had a volumetric productivity of 55 mg L-tyrosine/L/hr, representing 80% of the theoretical yield. Ajikumar et al. optimized a pathway for overproduction of taxadiene, a key taxol precursor³. The authors used a modular approach by separating the pathway into two operons, with one containing the methylerythritolphosphate pathway and the other containing genes encoding the downstream terpenoid-producing enzymes. The promoter strength in front of each operon was systematically varied and taxadiene product measured. Notably, the taxadiene production landscape was highly non-linear in response to operon expression.

Similar isogenic approaches can be used to engineer key rate-determining enzymes or transporters in implanted metabolic pathways. Zhang et al. used site-directed mutagenesis of active site residues of the enzymes KivD and LeuA⁴⁷. Fermentations of E. coli harboring pathways with different combinations of KivD/LeuA variants were tested for quantification of desired alcohol products. Leonard et al. generated combinatorial mutations in the downstream enzymes geranylgeranyl diphosphate synthase and levopimaradiene synthase to tune the selectivity and increase the productivity of levopimaradiene production in E. coli⁴⁸. The best strain had a maximum volumetric productivity of 7.3 mg levopimaradiene per L per h. Lee et al. improved xylose utilization in S. cerevisiae by directed evolution of xylose isomerase⁴⁹. After three rounds of error-prone PCR and screening they isolated a mutant with a 61-fold improvement in aerobic growth rate and an eight-fold improvement in ethanol production and xylose consumption. Screening pathway variants is not only limited to enzymes. Young et al. demonstrated the tunability of yeast sugar transporters through a combination of motif-based design and saturation mutagenesis⁵⁰. This approach was used to identify xylose-specific fungal molecular transporters, which when expressed improved xylose utilization by S. cerevisiae.

Computational Predictions Using Empirical Training Sets

Adjusting the right balance of enzyme specific activities within a pathway is crucial as the fitness cost of protein expression⁵¹, catabolism of pathway intermediates, and off-product reactions can all lower specific productivities. While there have been many admirable attempts to forward engineering biological systems and parts^{22,24,52–54} and analytical equations describing pathway flux have been formulated⁵⁵, tuning metabolic pathways is largely still an empirical exercise. Because of this, predictive computational models have been used to predict high productivity portions of sequence-flux space given sparse flux datasets resulting from testing isogenic cultures. Lee et al. used a linear regression model trained on empirical data to relate enzyme expression levels to product titers in a violacein biosynthetic pathway²⁹. This simple model could accurately predict promoter combinations resulting in the production of violacein or one of the three alternative products. Another approach to computationally model and improve pathway performance is to correlate targeted proteomics and metabolite data. George et al. generated isopentenol pathway variants with differing promoters, operon organization, and codon-usage³⁰. They then used HPLC and LC-MS to quantify glucose, organic acids, and pathway intermediates and used selected reaction monitoring mass spectrometry to quantify all proteins in their pathway. Spearman rank correlations were calculated from values of protein area and metabolite concentrations. Based on these relationships, individual variants were reconstructed and tested in time-course experiments to test model predictions. While this method may not capture complex regulatory interactions like feedback inhibition, other methods like ¹³C metabolic flux analysis studies are more than capable of doing so^56,57. In one example, Feng et al. tested different xylose reductase, xylitol dehydrogenase and xylulose kinase variants in a yeast xylose pathway and used ¹³C metabolic flux analysis to determine if the different cofactor requirements of the enzyme variants had any effect on growth and/or ethanol production⁵⁸. They found that production of ethanol was unaffected by the cofactor requirements of the xylose pathway. However, the cofactor balanced xylose pathway allowed growth under a wider variety of conditions. Farasat et al. developed a sequence-expression-activity mapping method to find optimal expression conditions with desired activity for a carotenoid biosynthetic pathway³¹. In a first step, an RBS calculator is used to make a library that spans a large range of protein expression space. Next, a subset of the library is tested for activity and used as a training set for a computational model. A new library is then constructed with targeted expression within a narrow window specified by the model. Zelcbuch et al. performed an iterative assembly of three fluorescent reporters, each with an associated RBS, into an operon⁵⁹. This initial search reduced the expression search space for a balanced astaxanthin pathway. In a clever approach, they were able to haplotype the individual non-local RBS sequences included within the operon by sequencing a downstream barcode built using iterative restriction and ligation steps.

Enabling DNA Construction Methods

New genetic modification methods like DNA Assembler⁶⁰, Golden Gate assembly⁶¹, Gibson cloning⁶², SLIC⁶³, site-specific recombination or VEGAS assembly⁶⁴ enable efficient construction of pathway variants with an array of different enzymes, promoters and RBS sequences. Smanski et al. utilized Gibson cloning⁶² and Golden Gate assembly⁶¹ to refactor the Klebsiella oxytoca nitrogen fixation gene cluster³² by systematically varying the expression levels of individual genes in the complete 16-gene pathway. Performance of their clusters was assessed by RNA-seq for expression levels and nitrogenase activity assays. The best of the 122 full-length pathways tested resulted in recovery of 57% of the wild-type activity. Layton and Trinh used Gibson cloning to make ester fermentative pathways in E. coli³³. Their modular design of their pathway allowed quick replacement of RBS and promoter sequences. Oliver et al. improved 2,3-butanediol production in cyanobacteria by using SLIC to swap different RBS sequences in front of each pathway enzyme³⁴. Colloms et al. used serine integrase site-specific recombination to rank gene order and RBS sites for a more efficient production of violacein and lycopene³⁵. Du et al. used S. cerevisiae native homologous recombination to swap promoters of various expression strength in front of relevant genes⁶⁵. This was used to improve xylose and cellobiose utilization pathways. Kim et al. used a similar approach to balance the flux of a xylose utilizing pathway for biofuel production³⁶. Importantly, the optimal pathway was strongly dependent on both the host genotype but also the sugar composition of the growth medium. Latimer et al. combinatorially tested promoters of the eight-gene pathway for xylose utilization in Saccharomyces cerevisiae²⁰. Library plasmids were made with Golden Gate assembly. Similar to results above, they found that the enrichment of specific yeast promoters in their library after selection was dependent on the number of genes expressed, the culture media conditions, and the cofactor dependence of the enzymes.

Alternative High-Throughput Screening Methods

Many of the above examples utilized medium-throughput plate-based screening or a growth based selection in order to sort variants. There have been recent developments to employ fluorescence-activated cell sorting (FACS) or microfluidic sorting technology in cases without an observable growth phenotype. For example, Wang et al. cultured xylose-consuming strains in droplets and microfluidic sorting based on the fluorescence of oxidized extracellular metabolites⁶⁶. Michener et al. utilized FACS to screen improved variants of caffeine demethylase using a designed RNA biosensor⁶⁷. The RNA biosensor is a combination of ribozyme and aptamer located in the 3′ UTR of a fluorescent reporter gene. When the aptamer is bound to a desired ligand, the ribozyme misfolds leading to lower RNA cleavage rates and increasing the fluorescent output. Tang et al. utilized FACS to screen for E. coli clones with enhanced triacetic acid lactone (TAL) production using a an engineered TAL fluorescent reporter⁶⁸. Jha et al. used a FACS screen to identify E. coli clones with increased enzymatic production of 3,4 dihydroxy benzoate⁶⁹. In these above examples, the limitation is developing a fluorescent reporter that is coupled to intracellular concentrations of a target metabolite.

Population-Based Measurements

One limitation of high-throughput screening is the inability to haplotype a unique pathway sequence to an output phenotype. Typically, only a few “winners” of the selection are sequenced. This is sub-optimal for two reasons. First, the winner variants depend strongly on the exact selection or screening conditions used, and so a selection must be repeated for each change of fermentation condition or host genotype. This limitation precludes use of multi-objective optimization (Pareto optimization) techniques into pathway redesign. Recent contributions from the Salis lab show the utility of Pareto optimization to improve translation elongation rates and mRNA stability of a single construct across multiple bacterial species^70,71. Secondly, high-throughput methods do not allow coverage of a complete sequence-flux space for even moderate-length pathways, and losing crucial genotypic information of the pathway makes it impossible to use the powerful computational analyses and prediction tools that have been demonstrated for low-throughput pathways. We envision population-based measurements that can more thoroughly search sequence-flux space and also identify Pareto optimal genotypes that are robust to different processing conditions (Fig. 2).

Figure 2.

Population-based measurements of pathways enable thorough search of flux space. (a) DNA barcoding methods allow long DNA constructs to be uniquely identified by short sequences. (b) Deep mutational scanning quantifies the enrichment of individual DNA variants after a selection. The enrichment relating the change in frequency of an individual variant can be related to specific productivity. (c) Comparing the fitness of individual variants between different selection conditions allows one to find Pareto optimal pathway sequences. This enables the identification of pathways supporting high specific productivity under robust processing conditions.

Recent advances in deep sequencing technology allow the ability to track tens of thousands of pathway variants in a high throughput screen. Most of such methods rely upon “barcoding” individual cells with a short unique identifier DNA sequence (Fig. 2a). A growth selection is performed, and these populations are deep sequenced at the barcode locus. The change in frequency of an individual barcode can be related to the fitness of that unique variant^72,73. While in principle such techniques could be used to track individual metabolic pathway variants, most demonstrations have been for studies on evolution. Smith et al. developed a barcode sequencing method (Bar-seq), which they validated by performing growth selections of a mixed culture containing yeast deletion strains⁷⁴. The barcode abundance after selection was determined for each deletion strain by deep sequencing the entire population. More recently, Levy et al. barcoded 500,000 lineages of Saccharomyces cerevisiae and used a growth selection to track time-dependent changes in fitness among the population⁷⁵. Chubiz et al. introduced FREQ-Seq, a method to barcode and determine allele frequencies from a mixed population⁷². FREQ-Seq was used to map seven variants of the enzyme Tet(X2), conferring tetracycline resistance, in ten different evolving populations⁷⁶.

Frequency analysis of variants within a population can be used to assess if a single variant improves, reduces, or has no effect on function. This approach has been used for evaluation of yeast translation initiation sites⁷⁷ and bacterial promoter strengths²¹ by coupling these upstream elements to fluorescent reporter proteins. Subsequently, populations are sorted by FACS. In fact, massively parallel sequence-function mapping is now commonplace in determining the sequence effects on function for proteins⁷⁸. For example, this methodology has been used to improve the affinity of engineered protein binders to Influenza⁷⁹. The question remains how to leverage impressive deep sequencing technology to improve implanted metabolic pathways.

In principle, high-throughput sequence-function mapping can be used to determine metabolic pathways supporting higher or lower flux, provided that it is coupled to a selectable phenotype like growth. A new approach called FluxScan was recently reported⁸⁰ that maps the sequence determinants of flux in living cells (Fig. 2b). First, a selection is designed to allow growth if and only if flux is routed through the implanted pathway. A mutational library is then created and transformed into the strain of interest. After a growth selection is performed for 4–10 generations, the entire population is deep sequenced and compared with the population before selection. The frequency change of each variant can be calculated and converted to a flux value. To demonstrate this method, Klesmith et al. determined the effect of flux for over 8,000 single-point mutants in a pyrolysis oil catabolic pathway⁸⁰. One designed pathway incorporating 15 beneficial mutations identified from FluxScan supported a 15-fold improvement in growth rate on levoglucosan, a chief pyrolysis oil constituent.

One significant technical challenge with FluxScan and related deep sequencing approaches is the inability to cover the complete length of metabolic pathways: current long read lengths of the Illumina platform are approximately 300 bp, whereas full operons can exceed 10 kb. One method to escape this limitation is to sequence small contiguous regions of sequence (a gene tile) able to fit on a single read. This “tiling” is then repeated along the length of the entire gene encoding sequence⁸¹. Other potential solutions to extend deep sequencing include coupling a pre-defined barcode sequence to a given pathway variant using clever DNA construction approaches^32,59 or to utilize the next generation of long-read, highly accurate haplotype sequencing technologies^82,83.

The ideal outcome of any experiment should be to find Pareto optimal pathway sequences that are robust and transferrable to any process condition. Single isogenic culturing conditions are not suited for this task as each pathway variant would individually have to be tested under each process condition to determine the resulting phenotype. Therefore, high-throughput population-based measurements are more capable to resolve the fitness of each sequence variant under each process condition (Fig. 2c) provided that they are performed under diverse conditions. Sequence variants from current high-throughput genomic methods that originate from different process conditions highlight this open problem. Gall et al. used the SCALEs method to map the gene expression in E. coli that conferred an advantage in the presence of 1-naphthol⁸⁴. They show that genes with enhanced expression depend on the type of culturing method used. Only 25% of clones that were reproducibly enriched in serial transfer cultures were similarly enriched in single batch cultures. Similarly, Warner et al. used TRMR and found differential gene expression that was dependent on the four different growth conditions under evaluation⁸⁵. While these examples highlight gene expression on the genomic scale, the problem also applies to expression and gene variants on the individual pathway level. An example of this is from the aforementioned FluxScan study where the enrichment of individual mutations from the enzyme levoglucosan kinase was strongly dependent on the biophysical properties of the starting enzyme variant from each selection⁸⁰. Being able to measure the pathway phenotype from different expression elements and individual gene variants under different growth conditions should allow the elucidation of pathway sequences that are optimal over a range of diverse conditions and determine why other sequences fail when these conditions change. If implemented, these high-throughput population-based measurements can help train computational models by providing empirical data and similarly computational models can help reduce the sequence search space for a more targeted population-based screen.

Outlook

In the near future we believe that robust, high performing pathways can be efficiently identified. New DNA assembly technologies allow for construction of large libraries of pathway variants covering a large range of protein expressions and activities. The number of unique pathway variants that can be made far exceeds that which can be accurately validated using existing technology. In this review, we have covered current methods to reduce the search space.

There is no general method that can assess any metabolic pathway, as there are limitations to each of the main approaches. There are two practical limitations that must be surmounted. First, the relationship between gene expression and pathway flux is highly non-linear. Second, a specific genotype may only support high productivity in a narrow range of process conditions. We suggest that marrying computational modeling with empirical datasets resulting from population-based measurements will allow a more efficient discovery of Pareto optimal gene encoding, expression, or regulatory sequences.