Advertisement
  • Loading metrics

The genetic basis for adaptation of model-designed syntrophic co-cultures

  • Colton J. Lloyd,

    Roles Conceptualization, Investigation, Methodology, Software, Writing – original draft, Writing – review & editing

    Affiliation Department of Bioengineering, University of California, San Diego, La Jolla, United States of America

  • Zachary A. King,

    Roles Conceptualization, Methodology, Software, Supervision, Writing – review & editing

    Affiliation Department of Bioengineering, University of California, San Diego, La Jolla, United States of America

  • Troy E. Sandberg,

    Roles Investigation, Writing – review & editing

    Affiliation Department of Bioengineering, University of California, San Diego, La Jolla, United States of America

  • Ying Hefner,

    Roles Investigation

    Affiliation Department of Bioengineering, University of California, San Diego, La Jolla, United States of America

  • Connor A. Olson,

    Roles Investigation

    Affiliation Department of Bioengineering, University of California, San Diego, La Jolla, United States of America

  • Patrick V. Phaneuf,

    Roles Software

    Affiliation Bioinformatics and Systems Biology Program, University of California, San Diego, La Jolla, United States of America

  • Edward J. O’Brien,

    Roles Methodology

    Affiliation Bioinformatics and Systems Biology Program, University of California, San Diego, La Jolla, United States of America

  • Jon G. Sanders,

    Roles Methodology

    Affiliations Department of Pediatrics, University of California, San Diego, La Jolla, United States of America, Cornell Institute of Host-Microbe Interactions and Disease, Cornell University, Ithaca, United States of America

  • Rodolfo A. Salido,

    Roles Methodology

    Affiliation Department of Pediatrics, University of California, San Diego, La Jolla, United States of America

  • Karenina Sanders,

    Roles Methodology

    Affiliation Department of Pediatrics, University of California, San Diego, La Jolla, United States of America

  • Caitriona Brennan,

    Roles Methodology

    Affiliation Department of Pediatrics, University of California, San Diego, La Jolla, United States of America

  • Gregory Humphrey,

    Roles Methodology

    Affiliation Department of Pediatrics, University of California, San Diego, La Jolla, United States of America

  • Rob Knight,

    Roles Methodology, Supervision

    Affiliations Department of Bioengineering, University of California, San Diego, La Jolla, United States of America, Department of Pediatrics, University of California, San Diego, La Jolla, United States of America, Center for Microbiome Innovation, University of California, San Diego, La Jolla, United States of America, Department of Computer Science and Engineering, University of California, San Diego, La Jolla, United States of America

  • Adam M. Feist

    Roles Conceptualization, Supervision

    afeist@ucsd.edu

    Affiliations Department of Bioengineering, University of California, San Diego, La Jolla, United States of America, Novo Nordisk Foundation Center for Biosustainability, Technical University of Denmark, Denmark

Abstract

Understanding the fundamental characteristics of microbial communities could have far reaching implications for human health and applied biotechnology. Despite this, much is still unknown regarding the genetic basis and evolutionary strategies underlying the formation of viable synthetic communities. By pairing auxotrophic mutants in co-culture, it has been demonstrated that viable nascent E. coli communities can be established where the mutant strains are metabolically coupled. A novel algorithm, OptAux, was constructed to design 61 unique multi-knockout E. coli auxotrophic strains that require significant metabolite uptake to grow. These predicted knockouts included a diverse set of novel non-specific auxotrophs that result from inhibition of major biosynthetic subsystems. Three OptAux predicted non-specific auxotrophic strains—with diverse metabolic deficiencies—were co-cultured with an L-histidine auxotroph and optimized via adaptive laboratory evolution (ALE). Time-course sequencing revealed the genetic changes employed by each strain to achieve higher community growth rates and provided insight into mechanisms for adapting to the syntrophic niche. A community model of metabolism and gene expression was utilized to predict the relative community composition and fundamental characteristics of the evolved communities. This work presents new insight into the genetic strategies underlying viable nascent community formation and a cutting-edge computational method to elucidate metabolic changes that empower the creation of cooperative communities.

Author summary

Many basic characteristics underlying the establishment of cooperative growth in bacterial communities have not been studied in detail. The presented work sought to understand the adaptation of syntrophic communities by first employing a new computational method to generate a comprehensive catalog of E. coli auxotrophic mutants. Many of the knockouts in the catalog had the predicted effect of disabling a major biosynthetic process. As a result, these strains were predicted to be capable of growing when supplemented with many different individual metabolites (i.e., a non-specific auxotroph), but the strains would require a high amount of metabolic cooperation to grow in community. Three such non-specific auxotroph mutants from this catalog were co-cultured with a proven auxotrophic partner in vivo and evolved via adaptive laboratory evolution. In order to successfully grow, each strain in co-culture had to evolve under a pressure to grow cooperatively in its new niche. The non-specific auxotrophs further had to adapt to significant homeostatic changes in cell’s metabolic state caused by knockouts in metabolic genes. The genomes of the successfully growing communities were sequenced, thus providing unique insights into the genetic changes accompanying the formation and optimization of the viable communities. A computational model was further developed to predict how finite protein availability, a fundamental constraint on cell metabolism, could impact the composition of the community (i.e., the relative abundances of each community member).

Introduction

Microbial communities are capable of accomplishing many intricate biological feats due to their ability to partition metabolic functions among community members. Therefore, these microbial consortia have the attractive potential to accomplish complex tasks more efficiently than a single wild-type or engineered microbial strain. Past applications include applying communities to aid in waste decomposition, fuel cell development, and the creation of biosensors [1]. In the field of metabolic engineering, microbial communities have now been engineered capable of enhancing product yield or improving process stability by partitioning catalytic functions among community members [28]. Beyond biotechnology applications, studying microbial communities also has important health implications. This includes providing a better understanding of the gut microbiome and how it is affected by diet and other factors [9,10]. For example, metabolic cross-feeding in communities has been shown to have a role in modulating the efficacy of antibiotics treatments [11]. New computational and experimental approaches to better understand the characteristics of viable microbial communities could therefore have far reaching implications.

Synthetic communities have been constructed to study their interactions and new metabolic capabilities. One such study encouraged synthetic symbiosis between E. coli strains by co-culturing an L-isoleucine auxotroph with a L-leucine auxotroph [12,13]. It was observed that the community was able to grow in glucose minimal media without amino acid supplementation due to amino acid cross-feeding between the mutant pairs. Mee et al. expanded upon this work by studying all possible binary pairs of 14 amino acid auxotrophs and developing methods to predict the results of combining the auxotrophic strains into 3-member, 13-member, and 14-member communities [14]. Similarly, Wintermute et al. observed community formation using a more diverse set of auxotrophs by co-culturing 46 conditionally lethal single gene knockouts from the E. coli Keio collection [15]. This work demonstrated that synthetic mutualism was possible in strains beyond amino acid auxotrophs [16]. These studies also demonstrated that new viable communities can be established in relatively short time frames (<4 days) by pairing auxotrophic strains.

In addition to establishing syntrophic growth, nascent auxotrophic communities can be optimized by adaptive laboratory evolution (ALE) [17]. Expanding upon the experimental work in Mee et al. [14], Zhang et al. performed ALE on one of the co-culture pairs: a L-lysine auxotroph paired with a L-leucine auxotroph [17]. Separate co-cultures evolved to growth rates 3-fold greater than the parent, which was accomplished, in part, by forming different auxotroph strain abundances within the community. Similarly, Marchal et al. evolved co-cultures of two E. coli amino acid auxotrophs and sequenced the endpoint strains. This data was leveraged to identify mutations hinting at changes in the spatial structure that occurred during the evolution [18]. Studies of evolved co-culture pairs composed of different microbial species have also used sequencing data and mutational analysis as a crucial component of interpreting adaptive strategies [19,20]. The success of the above work demonstrated that ALE can be used to optimize auxotrophic communities and that mutational data provide valuable insight into mechanisms underlying the evolved improvements in community growth rates.

Computational methods have been established to study the characteristics of microbial communities. These methods often apply genome-scale metabolic models (M-models) [2123]. Computational models have been created that use multicompartmental flux balance analysis (FBA) [2326], dynamic flux balance analysis (dFBA) [17,27], dFBA integrated with spatial diffusion of extracellular metabolites (COMETS) [28], and FBA with game theory [29]. Novel algorithms have also been developed to describe general community characteristics (OptCom [30]) and dynamics (d-OptCom [31]). These algorithms employ a bilevel linear programming problem to find the metabolic state that maximizes community biomass while also maximizing the biomass objectives of each individual species [32]. Numerous ecological models have also been formulated to describe community dynamics [3335].

Despite the significant advances made by the above modeling approaches, most methods were not intended to model suspension batch ALE experiments. For instance, ALE batch experiments in suspension assume growth in excess, well-mixed nutrients, thus negating the need for diffusion considerations (COMETS) or dynamic shifts in nutrient concentrations (dFBA). Also, in order for the strains to persist serial passage in an ALE experiment, it can be assumed that the cells in co-culture are growing, on average, at the same rate, thus negating the need for a bilevel growth objective that allows for varying growth rates of community members (OptCom). Additionally, given the growing appreciation for the role limited protein availability has on governing fundamental bacterial growth characteristics [36], it is likely that protein allocation plays a role in defining fundamental community characteristics as well. Therefore, there is a need for an applicable approach to model this experimental condition in a way that accounts for the protein cost of metabolism.

Here, we elucidate the genetic mechanisms underlying the formation of syntrophy between co-cultures of auxotrophic mutants containing diverse biosynthetic deficiencies. We first introduce the OptAux algorithm for designing auxotrophic strains that require high amounts of supplemented metabolites to grow (Fig 1A). The OptAux solutions provided a catalog of auxotrophic mutants representing a diverse set of metabolic deficiencies. From the catalog, four auxotrophic mutants were selected to co-culture and optimized via adaptive laboratory evolution (ALE) (Fig 1B). To increase the growth rate of the nascent co-culture communities, significant metabolic rewiring had to occur to allow the strains to cross-feed the high levels of the necessary metabolites. Some strains additionally had to adapt to marked changes in their homeostatic metabolic state, resulting from the inhibition of a major biosynthetic subsystem. The genetic basis accompanying this rewiring was assessed by analyzing the genetic changes (mutations and observed genome region duplications) over the course of the ALE. This mutational analysis further enabled predictions of primary metabolite cross-feeding and community composition.

thumbnail
Fig 1. Study overview.

(A) An algorithm was developed to de novo predict reaction deletions that will produce E. coli strains auxotrophic for a target metabolite. (B) From the set of auxotrophic strain designs, pairs were selected to determine whether they were capable of forming a viable syntrophic community. (C) The chosen co-cultures were both evolved via adaptive laboratory evolution and modeled using a genome-scale model of E. coli metabolism and expression (ME-model) [39,44]. The model predictions of fractional strain abundances and metabolite cross-feeding were compared to inferred results from the co-culture evolution experiments.

https://doi.org/10.1371/journal.pcbi.1006213.g001

To study the characteristics of the ALE-optimized communities, a community model of metabolism and expression (ME-model) was constructed [3739] (Fig 1C). Such a modeling approach was necessary since previous methods of genome-scale community modeling have focused on studying the metabolic flux throughout community members (using M-models) without consideration of the enzymatic cost of the proteins that drive these metabolic processes. As proteome optimization via niche partitioning and cell specialization is a driving factor of viable community formation in ecological systems [4043], it is essential to consider proteomic constraints when studying bacterial communities. To this end, community ME-models were utilized to interpret the nascent communities.

Results

OptAux development and simulation

The OptAux algorithm was designed to find metabolic reactions in E. coli that, when knocked out, will result in novel auxotrophies. This algorithm was implemented by selecting a metabolite of interest and applying OptAux to identify sets of reaction knockouts that will increase the uptake of the metabolite required for the cell to computationally grow (Fig 2A). OptAux was built by modifying an existing concept introduced for designing metabolite producing strains [45] which was later additionally implemented in a mixed-integer linear programming (MILP) algorithm (RobustKnock [46]). Three key modifications were made to derive OptAux from RobustKnock. First, the inner growth rate optimization was removed so that OptAux can be run at a predetermined growth rate (set_biomass constraint, Fig 2B). This ensures that OptAux designs computationally require the uptake of the target metabolite at all growth rates (Fig 2A, Figure A in S1 Appendix). Second, the objective coefficient was reversed in order to allow the algorithm to optimize for metabolite uptake as opposed to secretion. Third, a constraint was added to allow adjustments in the “specificity” of OptAux solutions (see Methods). This constraint allows the OptAux simulation to uptake any additional metabolite that can be consumed by the model (competing_metabolite_uptake_threshold constraint, Fig 2B). Without this constraint, many OptAux predicted designs have the potential to also grow in the presence of metabolites other than the target metabolite. For instance, it is possible that OptAux-predicted L-glutamate auxotroph mutants could alternatively grow when supplemented with L-glutamine or other metabolites as well. Therefore, “specificity”, in this case, refers to whether the mutant strain will be auxotrophic for a given metabolite in the presence of other metabolites. High specificity solutions are auxotrophic for only one metabolite, regardless of whether other metabolites are present. The implementation described above allowed OptAux to identify strain designs requiring the targeted metabolite at all growth rates with varying degrees of metabolite specificity.

thumbnail
Fig 2. OptAux design.

(A) OptAux was developed to maximize the minimum possible uptake of a target metabolite required for the model to grow. In other words, OptAux tries to increase the flux value at the intersection of the defined growth rate (set_biomass) and the minimum possible metabolite uptake flux (depicted with the red circle). Unlike algorithms such as OptKnock with tilting [47] and RobustKnock [48], the OptAux optimization occurs at a predetermined growth rate as opposed to imposing an inner growth rate optimization. This change was made to ensure that all OptAux designs will computationally require the uptake of a target metabolite at all growth rates, particularly low growth rates. The dotted lines show the required uptake for the metabolite with no genetic interventions. In this case, uptake of the target metabolite is not required at any growth rate. The solid black lines depicts the maximum and minimum uptake required for a particular metabolite in an OptAux designed strain. (B) The OptAux optimization problem. See Methods for further description of the algorithm and underlying logic.

https://doi.org/10.1371/journal.pcbi.1006213.g002

OptAux was utilized on the iJO1366 M-model of E. coli K-12 MG1655 [49,50] to comprehensively examine auxotrophic strain designs. OptAux was run with 1, 2, and 3 reaction knockouts for 285 metabolite uptake reactions using 4 different competing_metabolite_uptake_threshold values (S1 Data). Of the given solutions, 233 knockout sets were found to be capable of producing 61 unique strain auxotrophies. This set of strain designs provides an expansive look into the auxotrophies possible in the E. coli K-12 MG1655 metabolic network, which could be used to understand the possible niches that E. coli could inhabit in natural or synthetic communities [51].

OptAux solution characteristics.

The OptAux strain designs were broken into two major categories based on the number of individual metabolites that, when supplemented, can restore cell growth: 1) Essential Biomass Component Elimination Designs (EBC, Fig 3A) and 2) Major Subsystem Elimination Designs (MSE, Fig 3B). The EBC designs are characterized as auxotrophic strains with high metabolite specificity. They were broken into two subcategories: specific auxotrophs (only one metabolite can restore growth, Figure B in S1 Appendix) which consisted of 107 (20 unique) knockout sets and semi-specific auxotrophs (defined as strains in which less than 5 metabolites individually can restore growth, Figure B in S1 Appendix) which consisted of 67 (21 unique) knockout sets. The specific and semi-specific EBC designs were preferred at high competing_metabolite_uptake_threshold values.

thumbnail
Fig 3. OptAux solutions.

Two major solution types are possible depending on the parameters used when running OptAux. (A) Essential Biomass Component Elimination designs, like the ASNS1 and ASNS2 knockout shown, can grow only when one specific metabolite is supplemented. For the case shown, this metabolite is L-asparagine. (B) Alternatively, Major Subsystem Elimination designs have a set of alternative metabolites that can individually restore growth in these strains. Examples of these designs are shown for citric acid cycle knockouts sets. One specific three reaction knockout design (FUM, PPC, MALS) is shown in red dashed lines where four metabolites in the figure can individually rescue this auxotroph (marked with solid red circles). The metabolites that can restore growth for each of the knockout strain designs listed in the legend are indicated by the colored circles.

https://doi.org/10.1371/journal.pcbi.1006213.g003

There is notable overlap between OptAux predicted EBC designs (or those that are computationally identical), and known E. coli auxotrophic mutants [14,5263]. A summary of experimentally characterized OptAux designs is presented in Table A in S1 Appendix. Of note, there are 4 designs that were not found to be previously characterized in the scientific literature, and these present potential novel E. coli auxotrophs.

MSE designs were preferred at low competing_metabolite_uptake_threshold values and produce E. coli mutant strains with a diverse set of major metabolic deficiencies. These designs were defined as highly non-specific auxotrophic strains in which 5 or more metabolites could individually restore growth in the mutant strain. MSE designs consisted of the remaining 59 (20 unique) sets of knockouts. The MSE knockout strategy was often accomplished through knockouts that block metabolic entry points into key biosynthetic subsystems (Figure B in S1 Appendix). One such example of an MSE design is given in Fig 3B. Here a three reaction knockout design of the FUM, PPC, and MALS reactions can be rescued by one of the four compounds in the figure (i.e., citrate, L-malate, 2-oxoglutarate, or L-asparagine) at an average required uptake flux of 0.40 mmol gDW -1 hr -1 to grow at a rate of 0.1 hr -1. These rates are higher than the fluxes needed to rescue the EBC design in Fig 3A, which requires L-asparagine uptake of 0.024 mmol gDW -1 hr -1 at a rate of 0.1 hr -1. Another example of a novel MSE design was a glutamate synthase (GLUSy) and glutamate dehydrogenase (GLUDy) double knockout which effectively blocks the entry of nitrogen into amino acid biosynthesis by preventing its incorporation into 2-oxoglutarate to produce L-glutamate. This renders the cell unable to produce all amino acids, nucleotides, and several cofactors. In order to grow at a rate of 0.1 hr -1, this strain is computationally predicted to require one of 19 individual metabolites at an average uptake of 0.62 mmol gDW -1 hr -1 (S2 Data).

MSE designs are of particular interest as they are often unique, non-trivial, and have not been studied in the context of E. coli auxotrophies. However, some of the MSE single knockouts have been used for a large-scale study of auxotrophic co-culture short term growth [16]. Since these predicted MSE knockouts disrupt major metabolic flows in the cell’s biochemical network, they produce auxotrophies that require much larger amounts of metabolite supplementation to grow, compared to EBC designs (e.g., Figure C in S1 Appendix). To grow in co-culture, MSE E. coli mutants would require a pronounced metabolic rewiring and likely additional adaptation to a new homeostatic metabolic state, making them attractive to study from a microbial community perspective. Additionally, any strain paired with an MSE strain in co-culture would be required to provide a relatively high amount of the MSE strain’s auxotrophic metabolites to enable community growth.

Adaptive laboratory evolution of auxotrophic E. coli co-cultures

To demonstrate how the OptAux algorithm can be leveraged to design strains and co-culture communities, E. coli auxotrophic mutants were validated in the wet lab and evolved in co-culture. Three communities were tested, each consisting of pairwise combinations of four OptAux predicted auxotrophs. This included one EBC design, ΔhisD, which was validated as an L-histidine auxotroph, paired with each of three MSE designs, ΔpyrC, ΔgltAΔprpC, and ΔgdhAΔgltB. These three MSE strains had diverse metabolic deficiencies, including disruptions in pyrimidine synthesis, TCA cycle activity, and nitrogen assimilation into amino acids, respectively (Table B in S1 Appendix). The ΔpyrC mutant was computationally predicted to be capable of growing when supplemented with one of 20 metabolites in iJO1366, and the ΔgltAΔprpC and ΔgdhAΔgltB mutants were predicted to grow in the presence of 14 and 19 metabolites, respectively (S2 Data, Table D in S1 Appendix).

Four replicates of each co-culture were inoculated and initially exhibited low growth rates (< 0.1 hr -1), suggesting the strains initially showed minimal cooperativity or metabolic cross-feeding (Figure D in S1 Appendix). Following approximately 40 days of ALE, all 3 co-culture combinations had evolved to establish a viable syntrophic community, indicated by an increase in the co-culture growth rate. There was diversity in the endpoint batch growth rates among the independently evolved triplicates for each of the ΔhisD & ΔpyrC and the ΔhisD & ΔgdhAΔgltB co-cultures, with endpoint growth rates ranging from 0.09–0.15 hr -1 and 0.08–0.15 hr -1, respectively. The four successfully evolved independent replicates for the ΔhisD & ΔgltAΔprpC co-cultures also showed endpoint growth rate diversity ranging from 0.12–0.19 hr -1 (Table 1, Fig 4A). The relatively large range in endpoint growth rates for all co-cultures could suggest that a subset of replicates evolved to a less optimal state and thus could potentially be further improved if given more time to evolve. Alternatively, the slower growing co-cultures could have found a genetic state that resulted in a local maxima, rendering the co-culture less likely to increase its growth rate further.

thumbnail
Fig 4. Representative example of an adaptive laboratory evolution and its downstream analysis.

(A) E. coli co-cultures were evolved over a 40 day period and the growth rate was periodically measured. Over this time period the co-cultures evolved the capability to establish syntrophic growth, indicated by the improvement in community growth rate. (B) Each of the sampled co-cultures were sequenced at multiple points during the evolution. This information was used to predict the fractional strain abundances of each of the co-culture members (top panel, bars represent the computed fractional abundance of the strains in the legend). Sequencing data was also used to identify duplications in genome regions of the community members (middle panel) and infer causal mutations that improved community fitness (bottom panel). The complete set of ALE growth trajectories, inferred strain abundances, gene region duplications, and mutational analysis can be found in S1 Appendix, S3 Data, S4 Data, and Figs 57.

https://doi.org/10.1371/journal.pcbi.1006213.g004

thumbnail
Table 1. Starting and final growth rates, along with fractional strain abundance of the ΔhisD strain (by characteristic mutation), for each ALE lineage.

The cumulative number of cell division events that occurred throughout the experimental evolutions are also provided [64].

https://doi.org/10.1371/journal.pcbi.1006213.t001

To probe the adaptive strategies of the three co-culture pairs, the genomes of the populations were sequenced at several time points over the course of the 40 day evolution (Fig 4A). The sequencing data was used to identify genome region duplications and acquired mutations (Fig 4B), providing insight into the specific mechanisms employed by the co-cultures to establish cooperation.

The relative strain abundance of each mutant was tracked to observe the community composition throughout the course of the evolution. Each starting strain contained at least one unique characteristic mutation (Table C in S1 Appendix) that could act as a barcode to track the community composition (Fig 4B, Table 1). The breseq mutation identification software [65] was used to report the frequency of each of these characteristic mutations within a sequenced co-culture. The characteristic mutation frequency was then used to approximate the fraction of each strain within the co-culture population. This analysis showed that 2 of the 3 co-culture combinations maintained similar relative fractions of the two member strains, whereas one co-culture, ΔhisD & ΔpyrC, consistently maintained a relative ΔpyrC abundance of around three quarters of the total population (71–79%, Table 1). The strain’s prevalence in the community could potentially be overestimated if the strain’s characteristic mutations fell within duplicated genome regions. To account for this possibility, the relative abundance of each strain in the populations was additionally computed by comparing the read coverage of the knocked out genes for each mutant relative to the average read depth. This orthogonal method gave predictions consistent with those obtained using the characteristic mutation-based method (Figures E-F in S1 Appendix).

Following the evolutions it was confirmed that all collected ALE endpoint clones remained auxotrophic and had not evolved the ability to grow in glucose M9 minimal media. Given that only the large subunit (gltB) of glutamate synthase (catalyzes both glutamate synthase and glutamate dehydrogenase reactions, Table B in S1 Appendix) was knocked out, it was important to verify that the cell could not adapt to restore glutamate synthase functionality using only the small subunit (gltD) [66].

Mutations targeting metabolite uptake/secretion.

Several evolutionary strategies were observed in the mutations identified across the ten successfully evolved co-culture lineages (Tables E-G in S1 Appendix). One ubiquitous strategy across all three co-culture pairs, however, was to acquire mutations within or upstream of inner membrane transporter genes. For instance, numerous mutations were observed in every co-culture lineage in the hisJ ORF or upstream of the operon containing hisJ. This operon contains all four genes (hisJ, hisM, hisP, hisQ) composing the L-histidine ABC uptake complex, the primary mechanism for L-histidine uptake in E. coli K-12 MG1655 [67]. Seven mutations were found in the region directly upstream of the operon’s transcription start site (Fig 5). Two of the seven mutations were further observed in more than one co-culture pairing, with a SNP in one position (A->G, A->C, or A->T at 86 base pairs upstream of hisJ) appearing to be particularly beneficial as it was identified in the endpoint clone of every lineage except one (ALE #5). In three ALEs, a mutation was observed within the hisJ ORF that resulted in a substitution of the L-aspartate residue at the 183 position by glycine. Based on the protein structure, this substitution could disrupt two hydrogen bond interactions with the bound L-histidine ligand in the periplasm [68]. Alternatively, this mutation could function to modulate translation of the hisJ operon by altering its mRNA secondary structure. Further mutations were observed that could affect the binding of the ArgR repressor upstream of the hisJMPQ operon (Table E in S1 Appendix) or affect the activity of the ArgR protein itself (Table F in S1 Appendix). This included a 121 base pair deletion and a SNP in the ArgR repressor binding site upstream of hisJ (Fig 5). The mutation in the argR ORF consisted of a frameshift insertion early in the coding sequence and persisted throughout ALE #8, appearing in the ΔhisD endpoint clone (Table F in S1 Appendix). ArgR functions to repress L-arginine uptake and biosynthesis as well as repress the L-histidine ABC uptake complex [69] in response to elevated L-arginine concentrations. All of the above mutations could improve L-histidine uptake in the ΔhisD strains either by increasing the expression, improving the efficacy, or preventing ArgR mediated repression of the HisJMPQ ABC uptake system.

thumbnail
Fig 5. Mutations affecting inner membrane metabolite transport.

Mutations were observed that possibly affect the activity of four inner membrane transporters. A schematic of the function or putative function of each transporter is shown. Depicted below the schematics are the locations of the observed mutations on the operon encoding each of the enzymatic complexes. For example, all ten evolved ΔhisD strain endpoints possessed at least one mutation in or upstream of hisJ. This operon includes genes coding for HisJMPQ, the four subunits of an L-histidine ABC uptake system. A depiction of the activity of this complex is shown, in which energy from ATP hydrolysis is used to transport L-histidine into the cytosol from the periplasm. Mutations are indicated on the operon schematics if mutations appear at >10% frequency in more than one flask in an ALE lineage, and ALE numbers are in bold if the mutation appears in the endpoint clone. The mutations indicated with a dashed arrow occured in the ΔhisD strain and a solid arrow indicates they occured in ΔhisD strain’s partner MSE strain.

https://doi.org/10.1371/journal.pcbi.1006213.g005

Beyond improving the uptake of L-histidine in the ΔhisD strain, mutations were observed that could improve metabolite uptake in the partnering strain. For instance, in the ΔhisD & ΔgltAΔprpC co-culture, two of the evolutions acquired mutations in the kgtP ORF (a transporter of 2-oxoglutarate [70]) that were also present in the ΔgltAΔprpC endpoint clones. These mutations include a substitution of an L-proline residue with an L-glutamine at the 124 position and a substitution of a glycine residue with an L-alanine at the 143 position (Table E in S1 Appendix). These two substitutions occurred in the fourth transmembrane helix in the protein and a cytoplasmic region [71], respectively. These mutations could act to augment the activity of the transporter or modulate its expression by changing the mRNA secondary structure. The mutations further could complement the characteristic mutation upstream of the kgtP ORF observed in the starting clone of the ΔgltAΔprpC mutant (Table C in S1 Appendix). Both the accumulation of mutations associated with this transporter and the fact that the citrate synthase knockout mutant is computationally predicted to grow in the presence of 2-oxoglutarate suggest that ΔgltAΔprpC could be cross-fed 2-oxoglutarate in vivo when in co-culture (Table 2).

thumbnail
Table 2. Metabolite being cross-fed by the ΔhisD strain to its partner strain, as inferred from sequencing data.

https://doi.org/10.1371/journal.pcbi.1006213.t002

For the ΔhisD & ΔpyrC co-culture, mutations were consistently observed upstream of dctA that could function to better facilitate the uptake of a metabolite being cross-fed from the ΔhisD strain to the ΔpyrC strain. The three independently evolved lineages each acquired at least one mutation upstream of dctA, which were confirmed to be in all ΔpyrC endpoint clones (Table G in S1 Appendix). The gene product of dctA functions as a proton symporter that can uptake orotate, malate, citrate, and C4-dicarboxylic acids [72] (Fig 5). Model simulations of a ΔpyrC strain predicted that growth is possible with orotate supplementation, but not with any of the other metabolites known to be transported by the dctA gene product. Thus, it is possible these mutations could act to increase the activity of this transporter to allow the ΔpyrC strain to more efficiently uptake the orotate being cross-fed by the ΔhisD strain (Table 2).

Lastly, one lineage of the ΔhisD & ΔgdhAΔgltB co-culture acquired a SNP in the ygjI coding region and was present in the ΔhisD endpoint clone. This SNP resulted in a substitution of L-arginine for glycine at position 83, (Table F in S1 Appendix) within a periplasmic region and one residue prior to a transmembrane helix of the protein [73]. The function of this protein has not been experimentally confirmed, but based on sequence similarity, it is predicted to be a GABA:L-glutamate antiporter [74]. Given that this mutation was seen in the ΔhisD clone, it is possible that this mutation had the effect of increasing the strain’s secretion of 4-aminobutyrate (GABA) or L-glutamate by increasing the expression or modulating the activity of YgjI. Such a mutation could improve the community growth rate by facilitating the cross-feeding of either these metabolites to the ΔgdhAΔgltB strain since this strain is predicted to grow when supplemented with either GABA or L-glutamate (Table D in S1 Appendix).

Mutations targeting nitrogen regulation.

Knocking out enzymatic reactions in major biosynthetic pathways likely disrupts the homeostatic concentrations of key sensor metabolites, thus activating non beneficial stress responses (e.g., nutrient limited stress responses). The sequencing data was used to elucidate some of the adaptive mechanisms employed by the co-cultures following these pathway disruptions. For example, three frameshift deletions and a SNP resulting in a premature stop codon were observed early in the glnK ORF. These mutations were present in three ΔgltAΔprpC endpoint clones and one ΔhisD endpoint clone from the ΔhisD & ΔgltAΔprpC co-cultures (Fig 6B). GlnK along with GlnB are two nitrogen metabolism regulators with many overlapping functions. Both regulators are uridylated depending on the relative concentrations of 2-oxoglutarate, ATP, and L-glutamine. In conditions of high 2-oxoglutarate and ATP concentrations relative to L-glutamine concentrations, GlnK and GlnB are uridylated causing an increase in glutamine synthetase activity [75]. However, unlike GlnB, when GlnK is not uridylated it binds to the AmtB nitrogen uptake complex, thus reducing AmtB’s activity [76]. GlnK is also upregulated by GlnG of the nitrogen two-component regulatory system in the absence of nitrogen, unlike GlnB [77]. The citrate synthase knockout strain (ΔgltAΔprpC) in particular could see a disruption in the homeostatic concentrations of metabolites immediately downstream of the citrate synthase reaction, including 2-oxoglutarate and L-glutamine. This could impair the ability of the cell to respond to sensors of nitrogen excess or limitation and respond with the appropriate global regulatory changes. Removing the activity of this GlnK mediated response system would prevent any detrimental cellular responses (such as inhibition of the AmtB nitrogen uptake complex) due to atypical concentrations of the sensor metabolites within the co-culture strains. No mutations were observed in the alternative nitrogen regulator, GlnB, throughout any of the evolutions.

thumbnail
Fig 6. Mutations affecting nitrogen regulation.

Functions of the mutated genes are summarized, and the location of all mutations are shown on the operon below the schematic. Mutations are shown if they appear at >10% frequency in more than one flask in an ALE lineage, and ALE numbers are in bold if the mutation appears in the endpoint clone. The mutations indicated with a dashed arrow occured in the ΔhisD strain and a solid arrow if they occured in ΔhisD strain’s partner MSE strain. (A) Mutations were acquired within the open reading frame of both genes comprising the nitrogen sensing two-component regulatory system. Shown in the schematic is the regulatory cascade in which nitrogen concentration is sensed (via GlnK or GlnB) by GlnL. In response to low nitrogen availability GlnL is autophosphorylated resulting in a subsequent transfer of the phosphorus group to GlnG. Phosphorylated GlnG upregulates general functions associated with nitrogen starvation, including increasing GlnK expression [77]. (B) Further, mutations were observed in the ORF of GlnK, one of two nitrogen metabolism regulators, sharing most functions with GlnB. Both genes become uridylylated in response to high concentrations of 2-oxoglutarate and ATP and low concentrations of L-glutamine, which is an indication of nitrogen limitation. GlnK-UMP can activate GLNS deadenylation, thus increasing its activity. Unlike GlnB, when GlnK is in a deuridylylated state (indicative of high nitrogen availability) it can be sequestered by the AmtB ammonium transporter reducing the transporter’s activity [75].

https://doi.org/10.1371/journal.pcbi.1006213.g006

Mutations found in the ΔgdhAΔgltB strains imply a change in the activity of the two-component nitrogen regulatory system. The ΔgdhAΔgltB strain in all ΔhisD & ΔgdhAΔgltB lineages acquired mutations in the open reading frame of at least one gene in the two-component nitrogen regulator system, consisting of glnG (ntrC) and glnL (ntrB) (Fig 6A) [75]. Amino acid substitutions were observed in position 18, 86, and 105 of glnG corresponding to the response receiver domain of GlnG (based on protein families [78]), possibly augmenting GlnG’s ability to interact with GlnL. The endpoint clone of ALE #5 acquired an amino acid substitution of L-isoleucine to L-serine within a PAS domain of GlnL at position 12. This corresponds to the protein domain where regulatory ligands bind [79] suggesting this mutation could act to augment its activity in response to nitrogen availability. Like the citrate synthase knockout, the ΔgdhAΔgltB strain would likely experience a change in the homeostatic concentrations of metabolites used to sense nitrogen availability. Thus, it can be hypothesized that the mutations observed in the nitrogen two-component regulatory system act to augment the expression of nitrogen uptake and assimilation processes regulated by GlnGL.

Mutations were also observed targeting osmotic stress responses and nonspecific stress responses. These are summarized in the S1 Appendix.

Genome duplications complement sequence changes.

A complementary adaptive strategy for improving co-culture community growth was to acquire duplications in particular regions of the genome (Figures H-J in S1 Appendix). This evolutionary strategy possibly functioned in some cases to amplify expression of specific transporters to more efficiently uptake a metabolite that can rescue the strain’s auxotrophy (also observed in [80]). Alternatively, these duplications could function to provide genetic redundancy that increases the likelihood of acquiring mutations in the duplicated region [81,82]. For example, one of the three ΔhisD & ΔgdhAΔgltB lineages displayed clear increases in sequencing depth near positions 674–683 kbp and 1,391–1,402 kbp, with multiplicities exceeding 15. The former of these coverage peaks contains 9 genes, including the 4 genes composing the GltIJKL L-glutamate/L-aspartate ABC uptake system [83]. The latter peak consisted of 10 genes including the 4 genes in the abgRABT operon, which facilitates the uptake of p-aminobenzoyl-glutamate and its hydrolysis into glutamate and 4-aminobenzoate [84]. This suggests that either L-glutamate, L-aspartate, or p-aminobenzoyl-glutamate could be cross-fed to the ΔgdhAΔgltB strain in vivo. The abgRABT duplication, however, was depleted in favor of the gltIJKL duplication over the course of the evolution, suggesting L-glutamate or L-aspartate is the preferred cross-feeding metabolite over p-aminobenzoyl-glutamate (Fig 7, Table 2).

thumbnail
Fig 7. Duplication dynamics.

The top panel depicts the dynamics of high multiplicity duplications in two transport complexes throughout the course of ALE #5 of a ΔhisD & ΔgdhAΔgltB co-culture. A small region containing the abgT symporter of p-aminobenzoyl glutamate was duplicated early in the evolution, but later duplications in a region containing gltJ, along with the rest of the genes comprising the GltIKJL L-glutamate/L-aspartate ABC uptake system, became more prevalent. The bottom panel depicts the course of ALE #11, a ΔhisD & ΔgltAΔprpC co-culture that initially showed a broad ~1 Mbp duplication. By the end of the evolution either a nested duplication emerged in a small genome region containing hisJ, along with the rest of the HisJMPQ L-histidine ABC uptake system, or a significant subpopulation emerged containing this duplication.

https://doi.org/10.1371/journal.pcbi.1006213.g007

While the duplications mentioned above presented clear amplifications in targeted operons, some observed duplications consisted of 100,000s of basepairs and 100s of genes. Further, many of the duplications seen in the populations were not observed in the resequenced endpoint clones. Possible explanations for these observations can be found in the S1 Appendix.

Modeling community features of auxotroph communities

Community genome-scale models were applied to understand the basic characteristics of the co-culture communities generated in this study. Given the growing appreciation for the role of limited protein availability on governing many fundamental E. coli growth characteristics [36], community genome-scale models of metabolism and gene expression (ME-models) were utilized. A new computational approach was also developed, as a community modeling method did not exist that was suitable for studying co-cultures growing in an ALE experiment while also being amenable to ME-models (see Methods).

Using community M- and ME-models, the role of substrate and proteome limitations on basic community characteristics was assessed. To that end, both types of community models were constrained to uptake no more than 5 mmol gDWcommunity-1 hr -1 of glucose and simulated over a fractional ΔhisD strain abundance of 0 to 1 (Fig 8). The communities were allowed to cross-feed any metabolite that could restore growth in the partner strain (Table D in S1 Appendix). At this low glucose uptake rate the community ME-model was being simulated in the so-called substrate-limited region [37], meaning that the community growth rate was determined solely by the amount of substrate available. In this region the protein allocation constraints inherent in the ME-model were mostly inactive. In the substrate-limited region, the ME-model and M-model behaved similarly and predicted little change in the community growth rate regardless of the fractional abundance of the strains in co-culture. Alternatively, the community ME-model was again simulated, but with an unlimited amount of glucose available to the in silico community. These simulations therefore occurred in the proteome-limited region of the community ME-model, meaning that the growth rate was determined by limitations in the protein available to carry out their enzymatic functions. When simulating the community ME-model in the proteome-limited region, notable composition-dependent variation in the community growth rate was observed across all fractional strain abundances (Fig 8). Metabolite exchange for substrate- and proteome-limited ME-models was also observed (Figures M-N in S1 Appendix)

thumbnail
Fig 8. Comparison of community M- and ME-models.

The simulated growth rates for fractional strain abundances of ΔhisD ranging from 0 to 1. The top panel shows the community growth rate predictions of the community M-model and the community ME-model simulated in glucose-limited in silico conditions. The bottom panel shows growth rate predictions for the community ME-model simulations in glucose excess conditions. The arrows correspond to the fractional abundance that provided the highest computed community growth rate. The fractional abundances with growth rates greater than 95% of the maximum computed value were represented as a kernel density plot. The high density regions of the kernel density plot aligned well with the experimentally inferred community compositions, shown in the box plot.

https://doi.org/10.1371/journal.pcbi.1006213.g008

ME-model predictions are dependent on parameters that couple protein abundance to the flux values of the processes or reactions that they catalyze. These are called “keffs” and are analogous to the effective in vivo turnover rate of an enzyme. Obtaining these values on a genome-scale is a notoriously difficult problem [85], and no “gold standard” set of keffs currently exists. To account for uncertainty in these keff parameters, proteome limited community ME-model simulations were repeated using three different keff sets, including one set of naive values (“all keffs = 65”) and two sets derived using experimental data (“default model” [86] and “in vivo estimated keffs” [87,88]). All fractional abundance values within 95% of the maximum community growth rate were compiled and represented as a kernel density plot. The computed optimal community compositions (i.e., strain ratios that enabled the fastest computed community growth) showed relatively good agreement with the experimentally inferred community compositions (Fig 8). See the Methods for a description of the three keff sets.

The ME-modeling analysis suggested that it may be necessary to consider protein allocation when studying co-culture evolutions, therefore necessitating the use of resource allocation models, such as ME-models. The community ME-models thus were used to predict how the community composition could vary depending on basic characteristics of the co-cultures: 1) the identity of the metabolite that is cross-fed or 2) the enzyme efficiency of the community members. These simulations predicted that the metabolite being cross-fed within the community could have a sizeable impact on both the community composition and growth rate. This is particularly true for the ΔhisD & ΔgdhAΔgltB and ΔhisD & ΔgltAΔprpC simulations which showed that metabolite cross-feeding affected the growth rate and community compositions by as much as 50% (Fig 9A).

thumbnail
Fig 9. Community modeling.

Community ME-model-predicted growth rates computed with fractional strain abundances of ΔhisD ranging from 0 to 1. (A) The effect of metabolite cross-feeding on community structure. Each curve was computed after allowing each of the metabolites in the legend to exclusively be cross-fed to the MSE strain. Curves with identical computationally-predicted optimal strain abundances were grouped and given the same color. (B) The effect of varying the proteome efficiency of metabolite export on community structure (see Methods). The analysis was performed on models constrained to only cross-feed the metabolite that was considered most likely to be cross-fed to the ΔgltAΔprpC, ΔpyrC, and ΔgdhAΔgltB strains in vivo based on the sequencing data (2-oxoglutarate, orotate, and L-glutamate, respectively) (Table 2). (C) Box plots of experimentally inferred fractional strain abundances for each sample (bottom two rows, gray and dark blue) and the computationally-predicted optimal strain abundances following variation in the cross-feeding metabolite (top row, blue) and in strain proteome efficiency (second and third row, red and yellow).

https://doi.org/10.1371/journal.pcbi.1006213.g009

The strains growing in co-culture in vivo each undoubtedly differed in the protein cost required to synthesize the metabolite required by its partner strain. Therefore a proteome efficiency analysis (see Methods) was performed which showed that the computed optimal community compositions (the fractional strain abundance that gave the maximum community growth rate) of all three co-cultures were sensitive to the strain’s efficiency (Fig 9B). The computed optimal community composition was most sensitive when the ΔhisD strain’s metabolite export was less proteome efficient than its partner MSE strain. This observation is not surprising given that the ΔhisD strain must secrete metabolite(s) to the MSE strain at a much higher flux than the MSE strain to the ΔhisD strain. Therefore, a decrease in protein efficiency will have a larger impact on the ΔhisD strain. The community models also unintuitively predicted that, if the ΔhisD strain required a greater protein investment to produce the metabolite required by the partner strain (i.e., if the ΔhisD strain was less efficient than its partner), the abundances of the ΔhisD strain would actually increase in the community.

The optimal predicted community composition for the two above computational analyses shown in Fig 9A and 9B are summarized in Fig 9C. The figure shows general agreement between the computed optimal community compositions and the experimentally inferred community composition, even after varying key features of the community simulation (metabolite cross-feeding and protein efficiency). This suggests that community ME-models have the potential to be useful tools for understanding the behavior of simple communities. The same analysis was performed with the “in vivo estimated keffs” set of keffs and showed similar behavior (Figure O in S1 Appendix).

Discussion

This work provides genetic-level insight into the adaptation of model-designed nascent syntrophic communities growing cooperatively in suspension. This effort produced a novel algorithm, called OptAux, which was validated against historical auxotrophs and used to predict novel auxotrophic strain designs. OptAux-predicted designs with diverse metabolic deficiencies were co-cultured and community growth was optimized via adaptive laboratory evolution. Sequencing these co-cultures throughout the evolutions gave mutation and community composition information, thus providing insight into mechanisms of cellular cooperation. An additional modeling method was developed to interpret community features and demonstrated the importance of considering protein synthesis cost when studying cooperative communities in the utilized experimental conditions.

OptAux was demonstrated to be a useful tool for designing new types of cellular auxotrophies. Unlike many previously studied auxotrophies, OptAux enabled the prediction of auxotrophs stemming from a diverse set of major metabolic deficiencies. This included the prediction of 4 potential new essential biomass component elimination (EBC) designs and 20 unique major subsystem eliminations (MSE) designs. The OptAux-predicted MSE strains themselves could reveal further community insights if studied in co-culture. Co-cultures of two MSE strains would likely require a significant degree of metabolic rewiring in each strain to form a viable microbial community, thus probing the alternate evolutionary and cooperative paths such complex combinations could produce. OptAux is also suitable for predicting new auxotrophies in any organism outside of E. coli, provided the organism has an existing metabolic reconstruction [89].

Sequencing co-cultures throughout the course of the evolution experiments offered insight into the major adaptive mechanisms underlying the evolution of microbial cooperativity. The observed mutations indicated two major adaptive strategies employed by the strains in co-culture 1) mutating transporters, likely to improve uptake of auxotrophic metabolites (Fig 5) and 2) mutating to adapt to homeostatic changes as a result of metabolic disruptions upon imposing gene knockouts (Fig 6). The reported transporter mutations could prove useful for metabolic engineering applications, as optimizing the metabolite uptake characteristics of transporters can be an important component of improving the performance of engineered strains [90]. There, however, were no observed mutations, outside of mutations in a predicted GABA:L-glutamate antiporter in a ΔhisD strain, hinting at how the strains were capable of rewiring their intracellular metabolism to supply their partner strain with the required metabolite (i.e., no observed mutations associated with biosynthetic pathways). A future direction of this work could be to further evolve these strains to observe if new mutations appear to enhance metabolite rewiring. Alternatively, it is possible that the co-cultures grew by clumping and employing nanotube-mediated cross-feeding [91], which may be explored using microscopy.

Community ME-models were applied to understand the factors that drive community composition. This was the first community modeling effort to demonstrate the necessity of considering protein allocation when computationally studying community features. Interestingly, some of the studied co-cultures evolved to consistent community compositions that skewed away from a 50:50 strain ratio, a feature the community ME-models were often capable of capturing (Fig 8). Additionally, the community ME-models predicted that, if the ΔhisD strain became less protein efficient at producing the necessary cross-feeding metabolite, the optimal abundance of the ΔhisD strain in the co-culture would actually increase (Fig 9). Though unintuitive, this prediction is in agreement with a paradox highlighted in a previous computational study of community dynamics [92].

Despite the observed agreement between measured and computed optimal community compositions, this work highlighted the fact that there are a vast number of variables that could potentially influence basic features of simple communities. Experimentally assessing important features such as metabolite cross-feeding and community structure—as touched on here—on a large scale with many different cohorts and combinations is necessary to adequately understand the behavior of such bacterial communities. Model-driven design of communities and the use of community ME-models, however, present a more complete computational framework that can be leveraged as a tool to extract more knowledge from such experiments. Further, community ME-models offer a means to probe how factors outside of metabolism (e.g., translation efficiency and proteostasis) could affect community characteristics.

Materials and methods

Computational methods

All constraint-based modeling analyses were performed in Python using the COBRApy software package [93] and the iJO1366 metabolic model of E. coli K-12 MG1655 [49]. All optimizations were performed using the Gurobi (Gurobi Optimization, Inc., Houston, TX) mixed-integer linear programming (MILP) or linear programming (LP) solver. The community ME-models were solved using the qMINOS solver in quad precision [94,95]. All scripts and data used to create the presented results can be found at www.github.com/coltonlloyd/optaux.

OptAux algorithm formulation.

For the presented work it was necessary to employ an algorithm capable of finding reaction knockouts that would ensure the target metabolite is computationally essential in the in silico growth media for all feasible growth rates. To this end, a new algorithm was written as opposed to implementing a “reverse” version of RobustKnock (i.e., RobustKnock where the target objective is metabolite uptake instead of secretion). A “reverse” RobustKnock implementation would optimize the minimum required uptake of a metabolite at the maximum growth rate, thus leading to strain designs that must uptake a high amount of the target metabolite only when approaching the maximum growth rate (Figure A in S1 Appendix). To prevent this computational phenotype with OptAux, the inner problem optimizing for growth rate, which was utilized in RobustKnock, was removed. The growth rate was instead constrained to the set_biomass value, thus forcing the optimization to occur at a predefined growth rate. The constraint was implemented by setting the upper and lower bounds of the biomass objective function to set_biomass. Using relatively low set_biomass values with OptAux ensured the target metabolite would be computationally required for all feasible growth rates. For the simulations ran in this study (S1 Data), the set_biomass value was set to 0.1 hr -1.

An additional constraint was included in OptAux to represent additional metabolites present in the in silico media that could alternatively be used for growth, called the competing_metabolite_uptake_threshold. It was applied by finding all metabolites with exchange reactions and a default lower bound of 0 mmol gDW -1 hr -1 and increasing the bound to the competing_metabolite_uptake_threshold, thus allowing alternative metabolites in the in silico media to compete for uptake with the target metabolite. Increasing this threshold ultimately increases the specificity of the OptAux solution (i.e., whether other metabolites could potentially restore growth in addition to the target metabolite). In other words, if other metabolites were present in the in silico media, would the model still be auxotrophic for the target metabolite? If the strain would still be auxotrophic, it can be said to have high specificity; if the strain would not be auxotrophic, it can be said to be non-specific or semi-specific.

The resulting OptAux algorithm is a bilevel MILP (Fig 2B) that can be found at www.github.com/coltonlloyd/optaux.

OptAux simulations.

The OptAux algorithm was ran for all carbon containing metabolites with exchange reactions in iJO1366. The model’s default glucose M9 minimal in silico media was used for all optimizations with the maximum oxygen uptake set to 20 mmol gDW -1 hr -1. For each optimization the target metabolite was selected, and the maximum uptake of the metabolite was set to 10 mmol gDW -1 hr -1. The model was then reduced by performing flux variability analysis (FVA) on every reaction in the model and setting the upper and lower bounds of each reaction to the FVA results. If FVA computed that no flux could be carried through the reaction, then it was removed from the model. Additionally, reactions were excluded from knockout consideration if they met one of the following criteria: 1) it was an iJO1366 false positive when glucose is the primary carbon substrate [96] 2) it was essential in LB rich media [15] 3) its annotated subsystem was one of the following: Cell Envelope Biosynthesis, Exchange, Inorganic Ion Transport and Metabolism, Lipopolysaccharide Biosynthesis / Recycling, Murein Biosynthesis, Murein Recycling, Transport, Inner Membrane, Transport, Outer Membrane, Transport, Outer Membrane Porin, or tRNA Charging 4) it involved a metabolite with more than 10 carbons 5) it was a spontaneous reaction.

Identifying gene mutations and duplications.

The FASTQ data from the sequencing samples was filtered and trimmed using AfterQC version 0.9.6 [97]. The quality controlled reads were aligned to the genome sequence of E. coli K-12 BW25113 (CP009273.1) [98] using Bowtie2 version 2.3.0 [99]. Mutations were identified based on the aligned reads using breseq version 0.32.0b [65]. If the sample was of a co-culture population and not a clone, the predict polymorphism option was used with a frequency cutoff of 0.025. The output of the breseq mutation analysis for all samples can be found in S3 Data and on www.aledb.org [100].

Duplications were found by analyzing the BAM sequence alignment files output from Bowtie using the pysam Python package [101]. Pysam was used to compute the sequencing read depth at each DNA position within the genome sequence. For population samples, a cutoff of 1.25 x coverage fit mean (a measure of average read alignment coverage over the genome) was used. This relatively low threshold was used to account for the varying fractional abundances of the strains in community. A gene was flagged as duplicated in the sample if over 80% of the base pairs in the gene’s ORF had alignment coverage above the duplication threshold. Duplications found in starting strains were excluded from the duplication analysis. Further, the set of duplicated genes were grouped together if they were located next to each other on the genome. A new group was created if there existed more than five genes separating a duplicated gene from the next duplicated gene in the genome (S4 Data).

Aligned read coverage across the E. coli genome is noisy and therefore was filtered before plotting in order to observe its dominant features. This was accomplished by first splitting the coverage vector into 50,000 segments, such that each segment represented ~100 base pairs, and the average of the segments was found. Locally weighted scatterplot smoothing (LOWESS) was then applied to the array of concatenated segments using the statsmodel package in python [102]. For the smoothing, 0.5% of all of the segments was used when estimating each coverage value (y-value), and zero residual-based reweightings were performed. The remaining parameters were set to their default.

Calculating strain abundances from sequencing data.

The fractional abundances of the strains in co-culture were predicted using two features of the sequencing data obtained from each co-culture sample: 1) the frequency of characteristic mutations of each strain and 2) the read depth of the knocked out genes.

Each of the stains used in this study possessed a unique characteristic mutation (Table C in S1 Appendix), which could be used as a barcode to track the strain. The breseq mutation calling pipeline identified the characteristic mutations of each strain in co-culture and reported the frequency that the mutation was observed. This information was thus used to track the strain’s presence. For strains with two characteristic mutations (e.g., ΔhisD and ΔgdhAΔgltB) the reported frequency of the genes was averaged and used as a prediction of the relative abundance of that strain. One mutation in particular, an IS element insertion in yqiC, which is characteristic of the ΔhisD strain, was not detected in several samples when ΔhisD was in co-culture with ΔpyrC. This is likely due to the low abundance of the ΔhisD strain in that particular population. In those cases, the ΔhisD strain’s abundance was predicted using only the frequency of the lrhA/alaA intergenic SNP (Figure F in S1 Appendix). For one sample (A10 F23 I1 R1) the sequencing coverage was too low (~14.5) and the ΔgltAΔprpC characteristic mutation was not detected. Therefore no relative abundance was computed for this sample.

The second method for computing fractional strain abundances used the sequencing read alignment to compare the coverage of the deleted genes in each strain to the average coverage of the sample. As an example, for a strain paired with the ΔhisD strain, the average coverage of the base pairs in the hisD ORF divided by the average coverage for that sample, would give an approximation of its relative abundance in the population. As with the characteristic mutation approach, if the two genes were knocked out in the strain, the average coverage of the two genes was used to make the approximation (Figure E in S1 Appendix).

When reporting the relative abundance predictions (Figs 8 and 9), the computed abundances of each strain were normalized by the sum of the computed abundances of the two strains in co-culture. This ensured that the abundance predictions summed to one. Predictions made using the two described methods showed general agreement (Figure F in S1 Appendix).

Community modeling.

A community modeling approach was formulated that was amenable to ME-models and consistent with the characteristics of the ALE experimental design. The ALE experimental design applies a constant growth rate selection pressure by ensuring the cells are maintained in exponential growth phase in nutrient excess media conditions. A consequence of this experimental design when applied to co-culture systems is that the strains in co-culture must be growing at the same growth rate, on average. If this was not the case, one strain would be diluted from the culture or there would be dramatic fluctuations in the community composition, which is not the case (Fig 9C). Further, ALE experiments ensure that the culture is well mixed and grown in an excess of nutrients. These experimental conditions are not amenable to most existing community modeling methods. One modeling framework exists to study communities growing in steady state, called SteadyCom [23] (Figure L in S1 Appendix), though this method is not compatible with ME-models. This is due to the ME-model’s use of non-linear macromolecular coupling constraint expressions that are formulated as a function of growth rate. Therefore, the conversion to “aggregate biomass” flux used in the SteadyCom formulation cannot be translated directly to ME-models.

Given the above considerations, a multicompartment FBA approach, similar to community FBA [26] was used where the growth rates of the co-culture strains were constrained to be equal. The community model included one compartment for each of the two mutant strains in co-culture and a shared compartment where each of the strains could exchange metabolites. Further, the fluxes in and out of each strain’s compartment were scaled by the strain’s relative abundance to effectively mass balance the different model compartments (Figure K in S1 Appendix), thus allowing the relative abundance of each strain to be imposed as a parameter. For secretion, this was done by multiplying these exchange reactions as follows: and for uptake: where vsecrete is the secretion flux from strain 1 and has units of mmol gDWStrain1-1 hr -1 and XStrain1 is the fractional abundance of strain 1 with units of . Therefore, applying this coefficient to metaboliteShared gives fluxes in the shared compartment units of mmol gDWCommunity-1 hr -1. For the subsequent uptake of the shared metabolite by strain 2, the fractional abundance of strain 2 is applied giving units of mmol gDWStrain2−1 hr -1 (Figure K in S1 Appendix).

Using this community modeling approach, the fractional abundance of each strain in the co-culture was implemented as a parameter that could be varied from 0 to 1, which in turn had an impact on the optimal growth state of the community. All presented simulations were ran by optimizing the community growth rate for 10 values of XStrain1 (abundance of strain 1) ranging from 0.05 to 0.95. For XStrain1 values of 0 or 1 the community growth rate was assumed to be 0 hr -1 given that the co-culture mutants are auxotrophic and require the presence of both mutants to grow. The metabolites that were allowed to be cross-fed in simulation were limited to the set of metabolites that can computationally restore the growth of each auxotroph mutant (Table D in S1 Appendix).

For the community simulations, the iJL1678b [39] ME-model and iJO1366 [49] M-model of E. coli K-12 MG1655 were used. For proteome-limited ME-models simulations, the uptake of metabolites in the in silico glucose minimal growth media into the shared compartment was left unconstrained, as the ME-model is self-limiting [37]. For glucose-limited ME-model and M-model simulations, the maximum glucose uptake into the shared compartment was constrained to 5 mmol gDWCommunity-1 hr -1. The non-growth associated ATP maintenance and the growth associated ATP maintenance were set to the default parameter values in the model. For ME-model simulations, the RNA degradation constraints were removed to prevent high ATP costs at the low community growth rates. Since the newly formed communities are unoptimized and growing slowly, the ME-model’s unmodeled/unused protein fraction parameter was set to a higher value, 0.75, for proteome limited simulations (an unmodeled/unused protein fraction of 0.65 was imposed when the “in vivo estimated keffs” parameter set was used, since these keffs give a lower maximum growth rate than the other two keff vectors used) and the default value, 0.36, for glucose-limited simulations. If a metabolite had a reaction to import the metabolite across the inner membrane but no export reaction, a reaction to transport the metabolite from the cytosol to the periplasm was added to the model. For more on the ME-model parameters, refer to [39] and [37].

Three different sets of enzyme turnover rates (keffs) were used for the community ME-model simulations (Fig 8). The first set of keffs (“all keffs = 65”) was imposed by setting all keffs in iJL1678b-ME equal to 65 s-1. The next set of keff values (“default model”) used the default set of keff parameters included with iJL1678b-ME. Most of the metabolic keffs in this default set are determined by scaling a median keff value (65 s-1) by an estimation of the solvent accessible surface area of the enzyme complex that catalyzes the reaction (reference [37] for further description). The default keff parameters further included a set of 284 metabolic keffs derived using proteomics data and a computational method developed in Ebrahim et al. [86]. The last keff set (“in vivo estimated keffs”) included 234 keffs from Davidi et al. [87] that were estimated using model-computed fluxes and proteomics data. The keffs not estimated in Davidi et al. were imputed using the median estimated keff value from Davidi et al. (6.2 s-1). For all three keff sets, all non-metabolic processes were assigned a keff of 65 s-1.

Assessing the influence of metabolite cross-feeding on community composition was performed by restricting the simulation to cross-feed only one of the metabolites computationally predicted to restore growth in the MSE strain. In doing so, the identity of the metabolite being cross-fed could be related to the optimal community growth rate and structure.

To vary the proteome efficiency (keff) of secreting the cross-fed metabolites, first the exchange reactions into the shared compartment for all potential cross-feeding metabolites were constrained to zero, except the metabolite inferred from the experimental data (Table 2). Then the enzymatic efficiency of the outer membrane transport process of the inferred cross-feeding metabolite was altered in each strain. The outer membrane transport reactions for each inferred metabolite (i.e., HIStex, GLUtex, AKGtex, and OROTtex for L-histidine, L-glutamate, 2-oxoglutarate, and orotate, respectively) have multiple outer membrane porins capable of facilitating the transport process. To account for this, the keff kinetic parameter of each porin and reaction was changed by multiplying the default keff value by the appropriate multiplier. The COBRAme software was used for all ME-model computations [39].

Reproducibility.

All code and data necessary to reproduce the presented results can be found on GitHub at https://github.com/coltonlloyd/OptAux.

Experimental methods

E. coli strain construction.

All single gene knockouts used in this work were obtained from the Keio collection, a collection of all single gene knockouts in E. coli K-12 BW25113 [15]. To generate double gene knockout strains, the second knockout genes were identified from the Keio collection as donor strains, and their P1 phage lysates were generated for the transduction into the receiving single knockout strains. For instance, the ΔgltA or ΔgltB knockout strain was a donor strain and the ΔprpC or ΔgdhA knockout strain was a receiving strain (Table B in S1 Appendix), respectively. These four knockout strains were used for the construction of the double knockout strains, ΔgltAΔprpC and ΔgdhAΔgltB. Each mutant was confirmed not to grow in glucose M9 minimal media without supplementation of an auxotrophic metabolite predicted by the iJO1366 model.

Adaptive laboratory evolution.

Knockout mutants were each initially grown in lysogeny broth from a single colony, then washed 3 times and resuspended in M9-4g/L glucose medium. The washed cells from each knockout mutant preculture were then transferred to fresh M9-4g/L glucose medium and co-cultured with mutants from the partner strain. Cultures were initially inoculated with equal numbers of cells from the two relevant auxotrophs, then serially propagated (100 μL passage volume) in 15 mL (working volume) flasks of M9 minimal medium with 4 g/L glucose, kept at 37°C and well-mixed for full aeration. An automated system passed the cultures to fresh flasks once they had reached an OD600 of 0.3 (Tecan Sunrise plate reader, equivalent to an OD600 of ~1 on a traditional spectrophotometer with a 1 cm path length), a point at which nutrients were still in excess and exponential growth had not started to taper off. Four OD600 measurements were taken from each flask, and the slope of ln(OD600) vs. time determined the culture growth rates. The timescale of the evolution was reported using the cumulative number of cell divisions, as opposed to generations or days, as mutations occur primarily during cell division events [64].

Resequencing.

Co-culture population samples were collected at multiple midpoints throughout the ALE and sequenced. Additionally, the starting mutant strains and clones of both mutants isolated from the ALE endpoints were sequenced. The ΔhisD endpoint clone was unable to be isolated via colony selection for ALE #11. Genomic DNA of the co-culture populations and mutant clones was isolated using the Macherey-Nagel NucleoSpin tissue kit, following the manufacturer’s protocol for use with bacterial cells. The quality of isolated genomic DNA was assessed using Nanodrop UV absorbance ratios. DNA was quantified using the Qubit double-stranded DNA (dsDNA) high-sensitivity assay. Paired-end whole genome shotgun sequencing libraries were generated using KAPA HyperPlus kits and run on an Illumina MiSeq platform with a PE600v3 kit or an Illumina HiSeq 4000 with a PE-410-1001 kit for 150bp reads. DNA sequencing data from this study is available on the Sequence Read Archive database (accession no. SRP161177).

Supporting information

S1 Data. OptAux solutions.

Output of the OptAux algorithm ran for one, two, and three reaction knockouts on glucose minimal media for all carbon containing exchange metabolites. Four different competing metabolite uptake thresholds were used (0, 0.01, 0.1, 2).

https://doi.org/10.1371/journal.pcbi.1006213.s001

(XLSX)

S2 Data. Major subsystem elimination designs.

All MSE designs along with further information regarding the subsystems of the reaction knockouts and the metabolites that can restore growth in each design.

https://doi.org/10.1371/journal.pcbi.1006213.s002

(XLSX)

S3 Data. Mutations.

The breseq identified mutations for all samples collected in this work. Both the full output and a table with only mutations observed in the endpoint clones are provided.

https://doi.org/10.1371/journal.pcbi.1006213.s003

(XLSX)

S4 Data. Duplications.

Genes with read coverage meeting the duplication criteria. Separate spreadsheets are provided for all samples using the mutant pair, ale number, flask number, isolate number, and replicate number to identify each sample.

https://doi.org/10.1371/journal.pcbi.1006213.s004

(ZIP)

Acknowledgments

We thank Richard Szubin for help preparing samples for resequencing and thank Joshua Lerman and Justin Tan for informative discussions.

References

  1. 1. Rittmann BE, Hausner M, Löffler F, Love NG, Muyzer G, Okabe S, et al. A vista for microbial ecology and environmental biotechnology. Environ Sci Technol. 2006;40: 1096–1103. pmid:16572761
  2. 2. Minty JJ, Singer ME, Scholz SA, Bae C-H, Ahn J-H, Foster CE, et al. Design and characterization of synthetic fungal-bacterial consortia for direct production of isobutanol from cellulosic biomass. Proc Natl Acad Sci U S A. 2013;110: 14592–14597. pmid:23959872
  3. 3. Bernstein HC, Carlson RP. Microbial Consortia Engineering for Cellular Factories: in vitro to in silico systems. Comput Struct Biotechnol J. 2012;3: e201210017. pmid:24688677
  4. 4. Zuroff TR, Xiques SB, Curtis WR. Consortia-mediated bioprocessing of cellulose to ethanol with a symbiotic Clostridium phytofermentans/yeast co-culture. Biotechnol Biofuels. 2013;6: 59. pmid:23628342
  5. 5. Briones A, Raskin L. Diversity and dynamics of microbial communities in engineered environments and their implications for process stability. Curr Opin Biotechnol. 2003;14: 270–276. pmid:12849779
  6. 6. Zhang H, Pereira B, Li Z, Stephanopoulos G. Engineering Escherichia coli coculture systems for the production of biochemical products. Proc Natl Acad Sci U S A. 2015;112: 8266–8271. pmid:26111796
  7. 7. Zhou K, Qiao K, Edgar S, Stephanopoulos G. Distributing a metabolic pathway among a microbial consortium enhances production of natural products. Nat Biotechnol. 2015;33: 377–383. pmid:25558867
  8. 8. Saini M, Chen MH, Chiang C-J, Chao Y-P. Potential production platform of n-butanol in Escherichia coli. Metab Eng. 2015;27: 76–82. pmid:25461833
  9. 9. Flint HJ. The impact of nutrition on the human microbiome. Nutr Rev. 2012;70: S10–S13. pmid:22861801
  10. 10. Magnúsdóttir S, Heinken A, Kutt L, Ravcheev DA, Bauer E, Noronha A, et al. Generation of genome-scale metabolic reconstructions for 773 members of the human gut microbiota. Nat Biotechnol. 2017;35: 81–89. pmid:27893703
  11. 11. Adamowicz EM, Flynn J, Hunter RC, Harcombe WR. Cross-feeding modulates antibiotic tolerance in bacterial communities. ISME J. 2018; pmid:29991761
  12. 12. Hosoda K, Suzuki S, Yamauchi Y, Shiroguchi Y, Kashiwagi A, Ono N, et al. Cooperative adaptation to establishment of a synthetic bacterial mutualism. PLoS One. 2011;6: e17105. pmid:21359225
  13. 13. Hosoda K, Yomo T. Designing symbiosis. Bioeng Bugs. 2011;2: 338–341. pmid:22008942
  14. 14. Mee MT, Collins JJ, Church GM, Wang HH. Syntrophic exchange in synthetic microbial communities. Proc Natl Acad Sci U S A. 2014;111: E2149–56. pmid:24778240
  15. 15. Baba T, Ara T, Hasegawa M, Takai Y, Okumura Y, Baba M, et al. Construction of Escherichia coli K-12 in-frame, single-gene knockout mutants: the Keio collection. Mol Syst Biol. 2006;2: 2006.0008
  16. 16. Wintermute EH, Silver PA. Emergent cooperation in microbial metabolism. Mol Syst Biol. 2010;6: 407. pmid:20823845
  17. 17. Zhang X, Reed JL. Adaptive evolution of synthetic cooperating communities improves growth performance. PLoS One. 2014;9: e108297. pmid:25299364
  18. 18. Marchal M, Goldschmidt F, Derksen-Müller SN, Panke S, Ackermann M, Johnson DR. A passive mutualistic interaction promotes the evolution of spatial structure within microbial populations. BMC Evol Biol. 2017;17: 106. pmid:28438135
  19. 19. Summers ZM, Fogarty HE, Leang C, Franks AE, Malvankar NS, Lovley DR. Direct exchange of electrons within aggregates of an evolved syntrophic coculture of anaerobic bacteria. Science. 2010;330: 1413–1415. pmid:21127257
  20. 20. Hillesland KL, Lim S, Flowers JJ, Turkarslan S, Pinel N, Zane GM, et al. Erosion of functional independence early in the evolution of a microbial mutualism. Proc Natl Acad Sci U S A. 2014;111: 14822–14827. pmid:25267659
  21. 21. Zomorrodi AR, Segrè D. Synthetic Ecology of Microbes: Mathematical Models and Applications. J Mol Biol. 2016;428: 837–861. pmid:26522937
  22. 22. Perez-Garcia O, Lear G, Singhal N. Metabolic Network Modeling of Microbial Interactions in Natural and Engineered Environmental Systems. Front Microbiol. 2016;7: 673. pmid:27242701
  23. 23. Chan SHJ, Simons MN, Maranas CD. SteadyCom: Predicting microbial abundances while ensuring community stability. PLoS Comput Biol. 2017;13: e1005539. pmid:28505184
  24. 24. Klitgord N, Segrè D. Environments that induce synthetic microbial ecosystems. PLoS Comput Biol. 2010;6: e1001002. pmid:21124952
  25. 25. Freilich S, Zarecki R, Eilam O, Segal ES, Henry CS, Kupiec M, et al. Competitive and cooperative metabolic interactions in bacterial communities. Nat Commun. 2011;2: 589. pmid:22158444
  26. 26. Khandelwal RA, Olivier BG, Röling WFM, Teusink B, Bruggeman FJ. Community flux balance analysis for microbial consortia at balanced growth. PLoS One. 2013;8: e64567. pmid:23741341
  27. 27. Chiu H-C, Levy R, Borenstein E. Emergent biosynthetic capacity in simple microbial communities. PLoS Comput Biol. 2014;10: e1003695. pmid:24992662
  28. 28. Harcombe WR, Riehl WJ, Dukovski I, Granger BR, Betts A, Lang AH, et al. Metabolic resource allocation in individual microbes determines ecosystem interactions and spatial dynamics. Cell Rep. 2014;7: 1104–1115. pmid:24794435
  29. 29. Zomorrodi AR, Segrè D. Genome-driven evolutionary game theory helps understand the rise of metabolic interdependencies in microbial communities. Nat Commun. 2017;8: 1563. pmid:29146901
  30. 30. Zomorrodi AR, Maranas CD. OptCom: a multi-level optimization framework for the metabolic modeling and analysis of microbial communities. PLoS Comput Biol. 2012;8: e1002363. pmid:22319433
  31. 31. Zomorrodi AR, Islam MM, Maranas CD. d-OptCom: Dynamic multi-level and multi-objective metabolic modeling of microbial communities. ACS Synth Biol. 2014;3: 247–257. pmid:24742179
  32. 32. Feist AM, Palsson BO. The biomass objective function. Curr Opin Microbiol. 2010;13: 344–349. pmid:20430689
  33. 33. Biliouris K, Babson D, Schmidt-Dannert C, Kaznessis YN. Stochastic simulations of a synthetic bacteria-yeast ecosystem. BMC Syst Biol. 2012;6: 58. pmid:22672814
  34. 34. Oliveira NM, Niehus R, Foster KR. Evolutionary limits to cooperation in microbial communities. Proc Natl Acad Sci U S A. 2014;111: 17941–17946. pmid:25453102
  35. 35. Germerodt S, Bohl K, Lück A, Pande S, Schröter A, Kaleta C, et al. Pervasive Selection for Cooperative Cross-Feeding in Bacterial Communities. PLoS Comput Biol. 2016;12: e1004986. pmid:27314840
  36. 36. Basan M, Hui S, Okano H, Zhang Z, Shen Y, Williamson JR, et al. Overflow metabolism in Escherichia coli results from efficient proteome allocation. Nature. 2015;528: 99–104. pmid:26632588
  37. 37. O’Brien EJ, Lerman JA, Chang RL, Hyduke DR, Palsson BO. Genome-scale models of metabolism and gene expression extend and refine growth phenotype prediction. Mol Syst Biol. 2014;9: 693–693.
  38. 38. Lerman JA, Hyduke DR, Latif H, Portnoy VA, Lewis NE, Orth JD, et al. In silico method for modelling metabolism and gene product expression at genome scale. Nat Commun. 2012;3: 929. pmid:22760628
  39. 39. Lloyd CJ, Ebrahim A, Yang L, King ZA, Catoiu E, O’Brien EJ, et al. COBRAme: A computational framework for genome-scale models of metabolism and gene expression. PLoS Comput Biol. 2018;14: e1006302. pmid:29975681
  40. 40. Wilson M, Lindow SE. Coexistence among Epiphytic Bacterial Populations Mediated through Nutritional Resource Partitioning. Appl Environ Microbiol. 1994;60: 4468–4477. pmid:16349462
  41. 41. Zhao Q, Segre D, Paschalidisy IC. Optimal allocation of metabolic functions among organisms in a microbial ecosystem. 2016 IEEE 55th Conference on Decision and Control (CDC). 2016. https://doi.org/10.1109/cdc.2016.7799357
  42. 42. Teague BP, Weiss R. SYNTHETIC BIOLOGY. Synthetic communities, the sum of parts. Science. 2015;349: 924–925. pmid:26315419
  43. 43. Polz MF, Cordero OX. Bacterial evolution: Genomics of metabolic trade-offs. Nat Microbiol. 2016;1: 16181. pmid:27782136
  44. 44. O’Brien EJ, Lerman JA, Chang RL, Hyduke DR, Palsson BO. Genome-scale models of metabolism and gene expression extend and refine growth phenotype prediction. Mol Syst Biol. 2014;9: 693–693.
  45. 45. Feist AM, Zielinski DC, Orth JD, Schellenberger J, Herrgard MJ, Palsson BØ. Model-driven evaluation of the production potential for growth-coupled products of Escherichia coli. Metab Eng. 2010;12: 173–186. pmid:19840862
  46. 46. Tepper N, Shlomi T. Predicting metabolic engineering knockout strategies for chemical production: accounting for competing pathways. Bioinformatics. 2010;26: 536–543. pmid:20031969
  47. 47. Feist AM, Zielinski DC, Orth JD, Schellenberger J, Herrgard MJ, Palsson BØ. Model-driven evaluation of the production potential for growth-coupled products of Escherichia coli. Metab Eng. 2010;12: 173–186. pmid:19840862
  48. 48. Tepper N, Shlomi T. Predicting metabolic engineering knockout strategies for chemical production: accounting for competing pathways. Bioinformatics. 2010;26: 536–543. pmid:20031969
  49. 49. Orth JD, Conrad TM, Na J, Lerman JA, Nam H, Feist AM, et al. A comprehensive genome-scale reconstruction of Escherichia coli metabolism—2011. Mol Syst Biol. 2011;7: 535. pmid:21988831
  50. 50. Monk JM, Lloyd CJ, Brunk E, Mih N, Sastry A, King Z, et al. iML1515, a knowledgebase that computes Escherichia coli traits. Nat Biotechnol. 2017;35: 904–908. pmid:29020004
  51. 51. Zengler K, Zaramela LS. The social network of microorganisms—how auxotrophies shape complex communities. Nat Rev Microbiol. 2018; pmid:29599459
  52. 52. Fotheringham IG, Dacey SA, Taylor PP, Smith TJ, Hunter MG, Finlay ME, et al. The cloning and sequence analysis of the aspC and tyrB genes from Escherichia coli K12. Comparison of the primary structures of the aspartate aminotransferase and aromatic aminotransferase of E. coli with those of the pig aspartate aminotransferase isoenzymes. Biochem J. 1986;234: 593–604. pmid:3521591
  53. 53. Thèze J, Margarita D, Cohen GN, Borne F, Patte JC. Mapping of the structural genes of the three aspartokinases and of the two homoserine dehydrogenases of Escherichia coli K-12. J Bacteriol. 1974;117: 133–143. pmid:4148765
  54. 54. Glansdorff N. TOPOGRAPHY OF COTRANSDUCIBLE ARGININE MUTATIONS IN ESCHERICHIA COLI K-12. Genetics. 1965;51: 167–179. pmid:14292146
  55. 55. Jones-Mortimer MC. Positive control of sulphate reduction in Escherichia coli. Isolation, characterization and mapping oc cysteineless mutants of E. coli K12. Biochem J. 1968;110: 589–595. pmid:4882981
  56. 56. Sirko AE, Zatyka M, Hulanicka MD. Identification of the Escherichia coli cysM gene encoding O-acetylserine sulphydrylase B by cloning with mini-Mu-lac containing a plasmid replicon. J Gen Microbiol. 1987;133: 2719–2725. pmid:3329675
  57. 57. Somers JM, Amzallag A, Middleton RB. Genetic fine structure of the leucine operon of Escherichia coli K-12. J Bacteriol. 1973;113: 1268–1272. pmid:4570778
  58. 58. Wild J, Hennig J, Lobocka M, Walczak W, Kłopotowski T. Identification of the dadX gene coding for the predominant isozyme of alanine racemase in Escherichia coli K12. Mol Gen Genet. 1985;198: 315–322. pmid:3920477
  59. 59. Lee Y-J, Cho J-Y. Genetic manipulation of a primary metabolic pathway for L-ornithine production in Escherichia coli. Biotechnol Lett. 2006;28: 1849–1856. pmid:16933036
  60. 60. Felton J, Michaelis S, Wright A. Mutations in two unlinked genes are required to produce asparagine auxotrophy in Escherichia coli. J Bacteriol. 1980;142: 221–228. pmid:6102983
  61. 61. Vander Horn PB, Backstrom AD, Stewart V, Begley TP. Structural genes for thiamine biosynthetic enzymes (thiCEFGH) in Escherichia coli K-12. J Bacteriol. 1993;175: 982–992. pmid:8432721
  62. 62. Cronan JE Jr, Littel KJ, Jackowski S. Genetic and biochemical analyses of pantothenate biosynthesis in Escherichia coli and Salmonella typhimurium. J Bacteriol. 1982;149: 916–922. pmid:7037743
  63. 63. Yang Y, Tsui HC, Man TK, Winkler ME. Identification and function of the pdxY gene, which encodes a novel pyridoxal kinase involved in the salvage pathway of pyridoxal 5’-phosphate biosynthesis in Escherichia coli K-12. J Bacteriol. 1998;180: 1814–1821. pmid:9537380
  64. 64. Lee D-H, Feist AM, Barrett CL, Palsson BØ. Cumulative number of cell divisions as a meaningful timescale for adaptive laboratory evolution of Escherichia coli. PLoS One. 2011;6: e26172. pmid:22028828
  65. 65. Deatherage DE, Barrick JE. Identification of mutations in laboratory-evolved microbes from next-generation sequencing data using breseq. Methods Mol Biol. 2014;1151: 165–188. pmid:24838886
  66. 66. Mantsala P, Zalkin H. Active subunits of Escherichia coli glutamate synthase. J Bacteriol. 1976;126: 539–541. pmid:770440
  67. 67. Ardeshir F, Ames GF. Cloning of the histidine transport genes from Salmonella typhimurium and characterization of an analogous transport system in Escherichia coli. J Supramol Struct. 1980;13: 117–130. pmid:6449635
  68. 68. Yao N, Trakhanov S, Quiocho FA. Refined 1.89-A structure of the histidine-binding protein complexed with histidine and its relationship with many other active transport/chemosensory proteins. Biochemistry. 1994;33: 4769–4779. pmid:8161536
  69. 69. Caldara M, Charlier D, Cunin R. The arginine regulon of Escherichia coli: whole-system transcriptome analysis discovers new genes and provides an integrated view of arginine regulation. Microbiology. 2006;152: 3343–3354. pmid:17074904
  70. 70. Seol W, Shatkin AJ. Escherichia coli alpha-ketoglutarate permease is a constitutively expressed proton symporter. J Biol Chem. 1992;267: 6409–6413. pmid:1556144
  71. 71. Seol W, Shatkin AJ. Membrane topology model of Escherichia coli alpha-ketoglutarate permease by phoA fusion analysis. J Bacteriol. 1993;175: 565–567. pmid:8419306
  72. 72. Baker KE, Ditullio KP, Neuhard J, Kelln RA. Utilization of orotate as a pyrimidine source by Salmonella typhimurium and Escherichia coli requires the dicarboxylate transport protein encoded by dctA. J Bacteriol. 1996;178: 7099–7105. pmid:8955389
  73. 73. The UniProt Consortium. UniProt: the universal protein knowledgebase. Nucleic Acids Res. 2018; pmid:29425356
  74. 74. Riley M, Abe T, Arnaud MB, Berlyn MKB, Blattner FR, Chaudhuri RR, et al. Escherichia coli K-12: a cooperatively developed annotation snapshot—2005. Nucleic Acids Res. 2006;34: 1–9. pmid:16397293
  75. 75. van Heeswijk WC, Westerhoff HV, Boogerd FC. Nitrogen assimilation in Escherichia coli: putting molecular data into a systems perspective. Microbiol Mol Biol Rev. 2013;77: 628–695. pmid:24296575
  76. 76. Javelle A, Severi E, Thornton J, Merrick M. Ammonium sensing in Escherichia coli. Role of the ammonium transporter AmtB and AmtB-GlnK complex formation. J Biol Chem. 2004;279: 8530–8538. pmid:14668330
  77. 77. van Heeswijk WC, Hoving S, Molenaar D, Stegeman B, Kahn D, Westerhoff HV. An alternative PII protein in the regulation of glutamine synthetase in Escherichia coli. Mol Microbiol. 1996;21: 133–146. pmid:8843440
  78. 78. Finn RD, Coggill P, Eberhardt RY, Eddy SR, Mistry J, Mitchell AL, et al. The Pfam protein families database: towards a more sustainable future. Nucleic Acids Res. 2016;44: D279–85. pmid:26673716
  79. 79. Song Y, Peisach D, Pioszak AA, Xu Z, Ninfa AJ. Crystal structure of the C-terminal domain of the two-component system transmitter protein nitrogen regulator II (NRII; NtrB), regulator of nitrogen assimilation in Escherichia coli. Biochemistry. 2004;43: 6670–6678. pmid:15157101
  80. 80. Brown CJ, Todd KM, Rosenzweig RF. Multiple duplications of yeast hexose transport genes in response to selection in a glucose-limited environment. Mol Biol Evol. 1998;15: 931–942. pmid:9718721
  81. 81. Slack A, Thornton PC, Magner DB, Rosenberg SM, Hastings PJ. On the mechanism of gene amplification induced under stress in Escherichia coli. PLoS Genet. 2006;2: e48. pmid:16604155
  82. 82. Serres MH, Kerr ARW, McCormack TJ, Riley M. Evolution by leaps: gene duplication in bacteria. Biol Direct. 2009;4: 46. pmid:19930658
  83. 83. Wallace B, Yang YJ, Hong JS, Lum D. Cloning and sequencing of a gene encoding a glutamate and aspartate carrier of Escherichia coli K-12. J Bacteriol. 1990;172: 3214–3220. pmid:1971622
  84. 84. Carter EL, Jager L, Gardner L, Hall CC, Willis S, Green JM. Escherichia coli abg genes enable uptake and cleavage of the folate catabolite p-aminobenzoyl-glutamate. J Bacteriol. 2007;189: 3329–3334. pmid:17307853
  85. 85. Nilsson A, Nielsen J, Palsson BO. Metabolic Models of Protein Allocation Call for the Kinetome. Cell Syst. 2017;5: 538–541. pmid:29284126
  86. 86. Ebrahim A, Brunk E, Tan J, O’Brien EJ, Kim D, Szubin R, et al. Multi-omic data integration enables discovery of hidden biological regularities. Nat Commun. 2016;7: 13091. pmid:27782110
  87. 87. Davidi D, Noor E, Liebermeister W, Bar-Even A, Flamholz A, Tummler K, et al. Global characterization of in vivo enzyme catalytic rates and their correspondence to in vitro kcat measurements. Proc Natl Acad Sci U S A. 2016;113: 3401–3406. pmid:26951675
  88. 88. Heckmann D, Lloyd CJ, Mih N, Ha Y, Zielinski DC, Haiman ZB, et al. Machine learning applied to enzyme turnover numbers reveals protein structural correlates and improves metabolic models. Nat Commun. 2018;9: 5252. pmid:30531987
  89. 89. Feist AM, Herrgård MJ, Thiele I, Reed JL, Palsson BØ. Reconstruction of biochemical networks in microorganisms. Nat Rev Microbiol. 2009;7: 129–143. pmid:19116616
  90. 90. Kell DB, Swainston N, Pir P, Oliver SG. Membrane transporter engineering in industrial biotechnology and whole cell biocatalysis. Trends Biotechnol. 2015;33: 237–246. pmid:25746161
  91. 91. Shitut S, Ahsendorf T, Pande S, Egbert M, Kost C. Nanotube-mediated cross-feeding couples the metabolism of interacting bacterial cells [Internet]. 2017.
  92. 92. Kallus Y, Miller JH, Libby E. Paradoxes in leaky microbial trade. Nat Commun. Nature Publishing Group; 2017;8: 1361. pmid:29118345
  93. 93. Ebrahim A, Lerman JA, Palsson BO, Hyduke DR. COBRApy: COnstraints-Based Reconstruction and Analysis for Python. BMC Syst Biol. 2013;7: 74. pmid:23927696
  94. 94. Yang L, Ma D, Ebrahim A, Lloyd CJ, Saunders MA, Palsson BO. solveME: fast and reliable solution of nonlinear ME models. BMC Bioinformatics. bmcbioinformatics.biomedcentral. …; 2016;17: 391.
  95. 95. Ma D, Yang L, Fleming RMT, Thiele I, Palsson BO, Saunders MA. Reliable and efficient solution of genome-scale models of Metabolism and macromolecular Expression. Sci Rep. 2017;7: 40863. pmid:28098205
  96. 96. Orth JD, Palsson B. Gap-filling analysis of the iJO1366 Escherichia coli metabolic network reconstruction for discovery of metabolic functions. BMC Syst Biol. 2012;6: 30. pmid:22548736
  97. 97. Chen S, Huang T, Zhou Y, Han Y, Xu M, Gu J. AfterQC: automatic filtering, trimming, error removing and quality control for fastq data. BMC Bioinformatics. 2017;18: 80. pmid:28361673
  98. 98. Grenier F, Matteau D, Baby V, Rodrigue S. Complete Genome Sequence of Escherichia coli BW25113. Genome Announc. 2014;2. pmid:25323716
  99. 99. Langmead B, Salzberg SL. Fast gapped-read alignment with Bowtie 2. Nat Methods. 2012;9: 357–359. pmid:22388286
  100. 100. Phaneuf PV, Gosting D, Palsson BO, Feist AM. ALEdb 1.0: a database of mutations from adaptive laboratory evolution experimentation. Nucleic Acids Res. 2018; pmid:30357390
  101. 101. Li H, Handsaker B, Wysoker A, Fennell T, Ruan J, Homer N, et al. The Sequence Alignment/Map format and SAMtools. Bioinformatics. 2009;25: 2078–2079. pmid:19505943
  102. 102. Seabold S, Perktold J. Statsmodels: Econometric and statistical modeling with python. Proceedings of the 9th Python in Science Conference. SciPy society Austin; 2010. p. 61.