Toward Design-based Engineering of Industrial Microbes

Summary

Engineering industrial microbes has been hampered by incomplete knowledge of cell biology. Thus an iterative engineering cycle of modeling, implementation, and analysis has been used to increase knowledge of the underlying biology while achieving engineering goals. Recent advances in systems biology technologies have drastically improved the amount of information that can be collected in each iteration. As well, synthetic biology tools are melding modeling and molecular implementation. These advances promise to move microbial engineering from the iterative approach to a design-oriented paradigm, similar to electrical circuits and architectural design. Genome scale metabolic models, new tools for controlling expression, and integrated –omics analysis are described as key contributors in moving the field toward design-based engineering.

Introduction

Industrial microbiology has benefitted greatly from the introduction of recombinant DNA technologies, and the myriad of options that have hereby been enabled. These options have allowed a generation of engineers to splice DNA in many different ways, but even though recombinant DNA technology was developed in the early 1970s, many biotech challenges remain unsolved. While we knew how to splice DNA, the past ~30 years have been spent trying to figure out what DNA to splice. In other engineering disciplines, such as electrical, civil, and chemical engineering, design-based engineering has enabled the construction of engineering devices (bridges, electric circuits, etc.) directly from plans, because of a largely complete mathematical understanding of the physics that govern these devices. Microbial engineering is qualitatively different, in that the objective to date has been to (a) retrofit existing organisms to produce a product-of-interest using (b) knowledge and models that can be limited in their predictive capabilities.

As such, microbial engineering has traditionally followed a cyclic pattern of modeling, implementation, and analysis (Figure 1a). Typically, integrating available knowledge, e.g., the biochemical reactions known in an organism, into a mathematical model and use this for simulations can be used to develop a design strategy. Implementation of the strategies, such as extending a metabolic pathway to a new compound by introducing new enzymes, or knocking out competing pathways, is accomplished using recombinant DNA techniques consisting of plasmids or genomic alterations. Finally, the strain would be analyzed, sometimes only characterizing physiological properties and product yields. As models are limited and do not encompass all the activities of the cell, this process is inevitably iterative. Therefore, the knowledge gained from the previous analysis is used to refine the model, and the cycle begins again. Advances in Systems Biology leverage analysis and modeling to better define how a cell behaves in particular circumstances. Likewise, advances in Synthetic biology are allowing implementation of novel gene circuits that are guided by predictive models of each molecular ‘part.’ Combined, these recent advances in Systems Biology and Synthetic Biology are changing the efficiency of this engineering cycle toward the goal of transitioning from iterative engineering (Figure 1a) to a linear design-based engineering (Figure 1b) where modeling and implementation are adequately sophisticated to achieve the design goal without iteration. In this review we will highlight techniques and tools that are helping to move us toward design-based engineering of microbes for industrial purposes.

Figure 1. Transitioning from iterative to linear design-based engineering of industrial microbes.

(a) Iterative engineering - Genome scale models describe the cellular processes under desired condition. These models guide implementation of engineering strategies. These implementations are characterized by –omics and integrated analysis tools, which are used to revise the model and improve predictive capability. (b) Design-based Engineering – Models and implementation tools are reliable enough that the expected outcome is usually achieved, as in civil and electrical engineering. Transitioning from iterative cycle to a linear design-based engineering can be achieved by advance by better modeling and implementation aided by enhanced analysis.

Modeling

Models are a key asset to leverage in the design of microbes for a desired phenotype. Predictive models that describe relevant cellular processes can be used by engineers to rewire existing cellular processes toward a desired function. Genome-scale metabolic models (GSMMs) have been used extensively to predict changes in enzyme reaction rates (fluxes) in response to nutrient or genetic perturbations [1]. These models rely on (1) conservation of mass through the network including electrons and (2) cellular requirements for biomass to help predict how the cell will regulate metabolic pathways. Because these models predict changes in the metabolic network, they are of great use in predicting cellular alterations that will increase the production of a desired product [2]. These models can be used in combination with metabolic optimization algorithms, such as OptKnock or OptGene, to scan through all possible gene knock outs in the cell, and identify knockouts that are predicted to increase production [3,4]. Improvement in production titers for a large number of products, including: lactic acid, L-valine, L-threonine, other amino acids, hydrogen, vanillin (See ref [2] for details), lycopene[5], ethanol[6], and sequiterpenes [7] have been achieved using such models.

Stoichiometric models have been created for many industrially-relevant microbes, through curating enzyme biochemistry literature and using comparative genomics approaches for organisms that have little biochemical evidence. Genome-scale metabolic models exist for a range of industrial microbes. Table 1 gives information for published industrial microbes. Because of the rapid drop in sequencing costs, models for new organisms are being added often. Currently models are being developed for Penicillium chrysogenum and Pichia pastoris, which will be useful for antibiotic and protein production. Many of the published models can be found at the Biochemical, Genetic and Genomic (BiGG) database (http://bigg.ucsd.edu/) or the BioMet Toolbox (http://sysbio.se/BioMet/).

Table 1.

Genome-scale metabolic models of industrial microbes.

Organism	Reactions	Metabolites	Genes	Regulation	Thermodynamics	Reference(s)
Aspergillus nidulans	1,095	732	666			[55]
A. niger	2,168	1,045	871			[56]
A. oryzae	1,846	1,073	1,314			[53]
Bacillus subtilis	754	637	614			[57]
Corynebacterium glutamicum	495	408	411			[58]
Escherichia coli	2,077	1,039	1,260	✓	✓	[12]
Lactococcus lactis	621	422	358			[59]
Mannheimia succinicproducens	373	352	335			[60]
Saccharomyces cerevisiae	1,446	1,013	800			[61]
	1,489	972	805	✓		[10]
Streptomyces coelicolor	700	500	971			[62]

While there is an increasing availability of stoichiometric models for a range of organisms, for E. coli and S. cerevisiae, additional model features are being added to improve the models predictive capabilities. The regulatory structure that controls the expression of different metabolic enzymes is not explicitly captured in stoichiometric models. In E. coli, Boolean logic representations of transcriptional regulation have been able to predict binary expression changes in response to various stimuli [8]. Such models contain at least some regulatory information for approximately half the metabolic reaction and can be trained using high throughput phenotyping data and transcriptome measurements [9]. Similar efforts have been carried out in yeast, accounting for 55 transcription factors and nutrient conditions [10].

Another improvement has been identifying the reversibility/irreversibility of enzymatic reactions to constrain the model from predicting thermodynamically infeasible reactions. This is achieved by calculating the thermodynamic free energy of metabolites [11,12]. Reactions that have a significant drop in free energy in the forward direction are irreversible, while reactions without significant changes in free energy will be reversible, and subject to mass action.

Currently, these models are limited to quantify metabolic fluxes, which surely are of significant value for biochemical production. However, many processes that may be of interest to engineers are not described by such models. In protein production, for example, various sorting/quality control processes are controlled by interactions with other proteins, not transcriptionally controlled, nor can these processes be predicted by mass action. Other processes, such as DNA repair, cell division, signaling pathways, stress responses, and trafficking between compartments have inputs and outputs that may be partially transcriptional in nature, but have many other interactions that cannot be described by current FBA-type models (a notable exception is the description of the JAK-STAT signaling pathway by stoichiometric models [13]).

Protein-protein interaction networks provide a scaffold for models that describe non-metabolic processes in the cell and are currently a useful resource for data analysis (described below). Such networks do not have the quantitative predictive power as GSMMs, but serve as scaffolds to identify which proteins may be relevant without specific knowledge of the mechanism. Protein interaction networks are available for several industrial organisms, such as B. subtilis, E. coli, S. cerevisiae, and S. pombe, and are stored at BioGrid, a central repository for these and other types of interactions (www.thebiogrid.org) [14]. These databases are being significantly upgraded by various efforts and this will lead to a more complete mapping of many cellular processes and move toward the usefulness of protein-protein interaction data in predicting cellular processes.

Implementation

Predictive models can offer hypotheses for improving the desired phenotype. In general, implementation has consisted of (a) strong plasmid-based over-expression, (b) knockout, or (c) introduction of a heterologous gene. As we move forward with more detailed models, it will be desirable to move from the binary options of strong over-expression and deletion to fine-tuned alteration of constitutive expression levels, as well as sophisticated responsive promoters to adjust protein expression through the course of the fermentation (Figure 2).

Figure 2. Biological ‘parts’ for controlling synthesis / degradation rates in the central dogma paradigm.

Well-characterized biological components are necessary to control DNA, mRNA, and proteins in the cell. These biological components should be fine-tunable, independent of other implementations of the same control scheme, and robust in a variety of circumstances. The left column contains technologies that affect the synthesis rates, and the right column explains biological components that affect the degradation rates. Technologies are discussed in detail in the text. hRNA – Hammerhead RNA.

Well-characterized, robust expression systems will aid in this development. The Standard Biological Parts initiative (http://partsregistry.org/) is an open-source repository of biological “parts” (promoters, ribosome-binding sites, reporters, etc.) with input/output data compiled for convenient use of the part [15]. Abstracted parameters, such as polymerases per second (PoPs), are more useful in design than gfp fluorescence or β-galactosidase units that are commonly reported in research literature. This catalog relies on a cloning strategy called Biobricks which follows an ordered design experience that is less prone to user error than ad hoc cloning [16]. Biobricks has been been used by engineers with no cloning experience to design and implement cellular engineering strategies as can be seen by the student designs in the International Genetically Engineered Machine Competition (iGEM) [17].

New expression systems are being developed that expand the precision of expression levels and allow novel regulation. These approaches allow (a) fine tuned expression rather than just strong over-expression and (b) scalability, because the control systems can be used orthogonally, allowing extensive multiplexing rather than only a few options.

Constitutive expression over a range of expression levels has been achieved through mutagenesis of existing promoters, followed by characterization of the derived promoter library. This has been achieved in E. coli [18] and yeast [19] and the concept should be extendible to most relevant microorganisms. Expression strength can also be controlled by changing the copy number of tandem repeats on the chromosome using an approach called Chemically Induced Chromosomal Evolution (CIChE) [20]. CIChE has the added advantage of stable copy numbers, whereas plasmid copy number can fluctuate significantly.

Beyond constitutive expression, regulatable promoters allow engineers to dynamically control expression levels in responses to environmental changes, making new bioprocessing strategies possible. Similar to the mutation-based diversification used to create the constitutive promoters, Ellis, et. al. has created libraries of tetracycline and lactose responsive promoters in E. coli with altered response characteristics to their respective inducers [21]. The researchers then design a gene circuit and implement it with components of the library, a simple example of design-based engineering. Nevoigt et. al. has adjusted the oxygen sensitivity of the DAN1 promoter, allowing similar design possibilities [22].

De novo transcriptional regulatory networks can also be introduced into the cell, to co-regulate different groups of genes in parallel. This has been achieved by evolving zinc-finger transcription factors (TF) to recognize non-native promoter sequences [23]. Conceptually, an unlimited number of transcription factor/promoter combinations could be used independently to regulate genes. (for a more extensive review of these possibilities, see [24]). Combined with the mutation-based diversification of promoter characteristics, a range of promoter strengths could be made for a heterologous zinc-finger TF in both prokaryotes and eukaryotes.

Besides mRNA synthesis, other processes can affect expression strength. Hammerhead RNA (hRNA) detectors/regulators can control mRNA stability and translation efficiency (see [25] for an overview of RNA-based tools for controlling protein expression). These hRNAs can be evolved to detect a wide range of small molecule substrates, allowing different chemical inputs to affect protein synthesis. hRNAs are also scalable because different input/output hRNAs can operate without unintended cross-talk. Protein synthesis can also be controlled by altering the RBS and a recent model can be used to design specific translation activities [26].

Other means of controlling cellular function have moved beyond controlling RNA/protein abundance or are based on inheritance. To improve flux through a metabolic pathway, Dueber et. al. has developed a method to tether sequential heterologous enzymes in series [27]. By this, the pathways mimic the biological process of channeling providing increased local concentrations of intermediates and decreased average concentrations which (a) increases flux, (b) decreases toxicity of intermediates, and (c) avoids undesired side reactions. Light-controlled conditional protein interactions can allow direct control over the activity of enzymes, independent of transcription/translation [28,29]. Finally, genetic stability can be problematic where intermediates or products of a heterologous pathway are toxic to the cell. By switching from plasmids to tandem gene repeats on the chromosome, the CIChE method can extend genetic stability by tenfold [20].

Analysis

As models are not 100% predictive and molecular implementation may not perform as expected, the engineered cells must be analyzed after implementation. Over the past decade, technological advancements have enabled –omics measurement, the characterization of all instances of a class of molecules in a cell. In the wake of this massive amount of data that is generated, model-based analytic tools are allowing researchers to compress high dimensionality data to simpler, biologically-based insights, which can ultimately improve the model.

Transcriptomics examines the expression level of mRNAs at a specific physical condition. Various technologies have been developed to quantify the transcriptome, including hybridization (DNA microarray) and sequence-based approaches. The development of novel high-throughput sequencing methods has provided a new tool for both mapping and quantifying transcriptomes. This method, termed RNA-Seq (RNA sequencing), is being developed to profile transcriptomes and allows high throughput sequencing of cDNA in order to quantify mRNA in the cell [30]. Willenbrock et al. has shown that RNA-Seq correlates well with traditional microarray hybridization using synthetic RNA samples [31]. As well, the sensitivity in terms of reproducibility and relative quantification were equivalent. As microarray analysis requires an available genome sequence, custom chips for each organism, and complex statistical tools, RNA-seq, though a platform still under active development, overcomes these hybridization limitations, as well as enables analysis of sequence variants (for example SNPs [32]), and provides quantitative estimates of RNA abundance in the cell [33].

Proteomics attempts to quantify the level of all proteins in the cell. In the past, mRNA levels were used as a (sometimes poor) proxy for protein levels. Nowadays, methods combining mass spectrometry and liquid chromatography, either gel-based or non gel-based protein separation, are used to improve complex proteome identification. Since proteomics methods were developed in the middle of 1990’s, interactomics [34,35] and phosphoproteomics [36] have also become interesting for many research applications, characterizing the state of a protein. Now, various methods are developed for proteomics such as MALDI-TOF for peptide mass finger printing, and electrospray ionization (ESI) Fourier transform ion cyclotron resonance (FT-ICR) coupling with tandem mass spectrometry (MS/MS) for peptide identification. However, quantitative proteomic data must be validated by independent methods in order to overcome the heterogeneity of the biological samples [37]. Proteome analysis of Saccharomyces cerevisiae response to carbon and nitrogen limitation was done by using shotgun approach, multidimensional protein identification technology (MudPIT), in combination with the metabolic or in vitro labeling of proteins. The comparison of transcript and protein levels clearly shows that up-regulation in response to glucose limitation is transcriptionally controlled while up-regulation in response to nitrogen limitation is controlled at the post transcriptional level [38]. In a similar study the combination of transcriptome and proteome analysis by MudPIT allowed for global mapping of the regulatory network associated with the key protein kinase Snf1p in yeast [39].

Similar to proteomics, sophisticated mass spectrometry permits metabolomic studies. Metabolomics is achieved by quantifying cellular metabolites and systemically studying the unique chemical fingerprints that specific cellular processes leave behind. The methods coupling electrospray ionization with mass spectrometry is well developed and used for detecting the metabolomes in S. cerevisiae. With ESI-MS, up to 84% of the metabolites in S. cerevisiae can be identified [40]. A complete quantitative metabolomics workflow in a microplate format has recently been published [41]. Although the protocol is developed and validated for S. cerevisiae, this protocol can be adapted and applied for high throughput metabolomes studies in different organisms.

With this massive flow of complex data from mRNA, protein, and metabolite levels, integrative analysis is crucial and needed to prevent bottlenecks during the analytical step [42,43]. Integration comes from leveraging existing information about the cell through models (as described above) to interpret –omics data. At present, genotype-to-phetotype databases (G2P) are currently being improved using various integrative tools. The core modules for construction and improving genotype-to-phenotype databases include object model, exchange (data) format, ontology, globally unique identifier (GUID) and web service [44]. According to the availability of databases, a great number of mathematical models and integrative strategies have been developed and established as systems biology tools to integrate and visualize multidimensional data for examples PromoT [45], Metatool [46], PathSys and BiologicalNetworks [35] and YANAsquare [47]. Subnetwork analysis allows complex system with multistationarity such as cellular networks to be analyzed separately [48]. An algorithm to identify key metabolites and subnetwork structures in a biological response has been successfully developed by integrating transcriptome with topological information from genome-scale metabolic model without direct measurement of metabolite concentrations [49,50].

Many models are significantly revised after incorporating new biological information from the –omics analysis [51]. For example, a metabolic model of A. niger was significantly improved by combining metabolome and metabolic flux data with existing genome and literature data [52,53]. By integrating metabolic flux and transcriptome data with both interaction networks and stoichiometric models, a new approach to correlate mRNA and metabolic flux data was developed [34]. Through integrated data analysis, a better understanding of the biological system is possible, allowing more realistic models to be developed for bio-based industries.

Conclusions & Future Directions

As synthetic genomes are quickly becoming a reality [54], we are then faced with an analogous question as 30 years ago. We know how to synthesize DNA, but what DNA should we synthesize? These research efforts mentioned here are qualitatively changing the nature of the cyclic engineering paradigm for industrial microbes. Thanks to increased analysis capability, much more information is gathered in each iteration. As well, with better models and molecular tools, better hypotheses and more controlled implementation is now possible.

Further developments in models that capture the change in metabolite concentrations will be useful in predicting flux distribution as well as potentially toxic levels of a particular metabolite. To improve molecular implementation, a larger catalog of well characterized parts will need to be available to the research community. The standard biological parts initiative has, thus far, been an open source initiative. With suitable intellectual property security, private biological “parts” companies could do rigorous characterization and sell “components” to microbial engineers, significantly accelerating the accumulation of available parts. And finally, the use of –omics data to create phenotype-predictive models of protein-protein interactions would significantly extend the biological processes that could be engineered. In the end, the far, far end, experimental analysis may become an academic exercise for undergraduate, as is the state in electrical circuit design. At this point, we will have reached a paradigm of design-based engineering of industrial microbes.

Figure 3. The conceptual work flow of integrated analysis of –omics data to formalize genotype-to-phenotype database.

Advances in -omics measurement lead to the accumulation of high throughput -omics data. With systematic and integrative analysis tools, genotype-to-phenotype database are currently being developed under the core modules outlined in this figure. Standardized object models and data formats facilitate the exchange of information between different data systems and are essential for unambiguous transmission of data between computers. A well defined ontology enables the representation of domain-specific knowledge, as well as, the relationship between those domains and leads to powerful database searching. GUIDs solve data integration problems that result from ambiguity in name, or identity, of biological concepts and objects. Standard protocols for web service afford machine-to-machine interaction over the internet and simplify the task of exploring distributed data, forming the basis of the service-oriented architecture (SOA) [44].