Predicting outcomes of steadystate ^{13}C isotope tracing experiments using Monte Carlo sampling
 Jan Schellenberger^{1},
 Daniel C Zielinski^{2},
 Wing Choi^{2},
 Sunthosh Madireddi^{2},
 Vasiliy Portnoy^{2},
 David A Scott^{3},
 Jennifer L Reed^{4},
 Andrei L Osterman^{3} and
 Bernhard ∅ Palsson^{2}Email author
https://doi.org/10.1186/1752050969
© Schellenberger et al. ; licensee BioMed Central Ltd. 2012
Received: 4 August 2011
Accepted: 30 January 2012
Published: 30 January 2012
Abstract
Background
Carbon13 (^{13}C) analysis is a commonly used method for estimating reaction rates in biochemical networks. The choice of carbon labeling pattern is an important consideration when designing these experiments. We present a novel Monte Carlo algorithm for finding the optimal substrate input label for a particular experimental objective (flux or flux ratio). Unlike previous work, this method does not require assumption of the flux distribution beforehand.
Results
Using a large E. coli isotopomer model, different commercially available substrate labeling patterns were tested computationally for their ability to determine reaction fluxes. The choice of optimal labeled substrate was found to be dependent upon the desired experimental objective. Many commercially available labels are predicted to be outperformed by complex labeling patterns. Based on Monte Carlo Sampling, the dimensionality of experimental data was found to be considerably less than anticipated, suggesting that effectiveness of ^{13}C experiments for determining reaction fluxes across a largescale metabolic network is less than previously believed.
Conclusions
While ^{13}C analysis is a useful tool in systems biology, high redundancy in measurements limits the information that can be obtained from each experiment. It is however possible to compute potential limitations before an experiment is run and predict whether, and to what degree, the rate of each reaction can be resolved.
Background
There are several distinct sources of variability in a ^{13}C experiment that limit the confidence with which particular reactions can be determined. First, due to experimental accuracy limitations and biological variability, uncertainty arises in the experimentally measured MDV. Second, due to alternate pathways present in metabolic networks, the mass balance equations underlying a metabolic steadystate are significantly underdetermined [9]. While the full network flux distribution may not be resolved at high confidence by a given experiment, certain labeling patterns may resolve fluxes through certain pathways with greater confidence than other labeling patterns, as has previously been shown [4, 10].
For a given ncarbon compound, there are 2 ^{ n } possible ^{13}C labeling states (as well as mixtures), and the choice of label is known to affect the ability to determine reactions fluxes [10]. As ^{13}C methods are based upon computational modeling of isotopomer distributions, it is possible to computationally optimize the choice of substrate labeling pattern to enhance the information gained from an experiment. There are two primary motivations that drive such an endeavor. First, ^{13}C experiments are expensive, so choosing the best experiment a priori is desirable. Second, we can assess the capability of the steadystate ^{13}C labeling approach towards determining reaction fluxes in an unbiased manner. The issue of optimization of ^{13}C labeling experiments has been addressed in the literature [4, 10, 11]. However, the use of flux sampling for optimal isotopomer experiment prediction has not been explored previously, and this approach presents several unique advantages over previous methods.
We describe a Monte Carlo samplingbased method for choosing the optimal substrate label, based upon the ConstraintBased Reconstruction and Analysis (COBRA) computational platform [12, 13]. COBRA methods use manuallycurated biochemical network reconstructions of known reaction stoichiometries and measurable nutrient uptake and secretion rates to define feasible ranges for internal reaction fluxes. Many of these reconstructions have been generated [14] and the procedure is wellestablished [13, 15]. These models can be used for methods such as computing growth rates [16, 17], predicting the effects of gene knockouts [16, 18, 19], predicting the endpoint of adaptive evolutions [20], and designing strains for industrial production [21, 22]. A review of these methods can be found here [12, 13, 23]. Monte Carlo sampling of constraintbased metabolic models can be used to generate sets of biochemically feasible flux distributions that obey measured uptake and secretion rate constraints [24]. IDVs generated from these flux distributions in an isotopomer model can then be compared against simulated ^{13}C data to evaluate the ability of the experiment to determine reaction fluxes. Monte Carlo sampling takes advantage of the speed with which IDVs can be simulated from putative flux distributions, making this approach suitable for largescale analysis of in silico experiments.
A Monte Carlo sampling approach was implemented using a newly developed isotopomer model to evaluate the efficiency of different carbon labeling patterns toward determining reaction fluxes in E. coli. The dimensionality of simulated ^{13}C data was calculated using singular value decomposition (SVD) for different substrate labeling patterns and compared to the number of undetermined dimensions in the network. ^{13}C experiments were performed for three substrate labeling patterns to validate the prior theoretical analysis. The methods developed represent a flexible computational analysis that can be applied to various biological systems and experimental setups to estimate, a priori, the efficiency of isotopomer experiments in determining reaction fluxes.
Results and Discussion
Expanded Isotopomer Model
An isotopomer model was constructed in two phases. First, a central metabolic isotopomer model that accounts for 85 reactions including glycolysis, the TCA cycle, the pentose phosphate pathway, oxidative phosphorylation, pyruvate metabolism, and anaplerotic reactions was derived from the iJR904 E. coli reconstruction [16]. This initial model was equivalent in reaction content to commonly used isotopomer models for E. coli [25, 26].
An expanded model was then constructed that includes both central and biosynthetic pathways. The iMC1010 metabolic network [19] was evaluated to determine which reactions can sustain nonzero fluxes during growth on glucose, acetate, or lactate when only certain byproducts are allowed to be secreted (acetate, formate, Dlactate, pyruvate, succinate, glycerol, CO_{2}, and ethanol). Blocked reactions, which must have zero net flux at steady state, were subsequently omitted from consideration. Groups of reactions that could be merged together without affecting model results (e.g. linear pathways) were combined in order to reduce the number of variables. Large sets of biosynthesis reactions that produce phospholipids, nucleotides, cofactors were also combined, since there no experimental measurements existed for these highcarbon metabolites. However, byproducts resulting from highcarbon metabolite production (e.g. CO_{2}, formate, succinate, fumarate, and pyruvate) that could enter back into the metabolic network were tracked. Of the original 932 reactions in the complete metabolic iMC1010 network, nearly a third were represented in the biosynthetic isotopomer model, either individually or as grouped reactions.
The final isotopomer model accounts for a total of 313 irreversible reactions, including 278 which track carbon. Inclusion of these additional pathways is likely important for accurate assessment of the fluxresolving power of ^{13}C experiments both within and beyond central metabolism [7]. A complete listing of the reactions and metabolites in the biosynthetic network can be found in the Additional File 1.
Monte Carlo Sampling Approach
To compute possible flux distributions of the E. coli model, the network was sampled using a Markov Chain, Monte Carlo (MCMC) sampling algorithm (see Methods). The steadystate mass balance and uptake rate constraints for the metabolic network create a convex hyperspace that contains all biochemically feasible steadystate flux distributions [27]. Monte Carlo sampling generates a set of flux distributions that are spread uniformly throughout the feasible space. The inclusion of ^{13}C experimental data reduces the feasible space in which the true flux state must lie by requiring that the IDV calculated from the putative flux distribution must match the experimental data within error. While the space of feasible flux distributions depends only on reaction stoichiometry, the space of resulting simulated IDVs differs depending on the input substrate labeling pattern. Hence, different labeling patterns can have differing abilities to resolve each reaction flux.
Generating and Evaluating ^{13} C Experimental Hypotheses
An experimental hypothesis is defined as a partition of the sampled flux distribution set. While many possible hypotheses could be considered, two rational hypotheses were studied. The first case attempts to elucidate whether a reaction has high or low flux (hilo). The solution space is partitioned into all points with v_{ j } > threshold versus v_{ j } < threshold. A different hypothesis is generated for each reaction j. The threshold was chosen to be the median of all v_{ j } so that half of all points would be in each of the two partitions. The second set of hypotheses tested consisted of biologically relevant flux ratios. For each point the ratio of two reactions, v_{ i } /v_{ j } , was determined to be above or below some threshold that formed a partition.
Using this approach, candidate flux states were sampled uniformly and experimental hypotheses tested. Zscores were calculated for the hilo hypothesis corresponding to 1) individual reactions 2) reaction ratios and 3) two 'random' reactions or ratios. Random hypotheses were tested to estimate the level of noise associated with the set of flux distributions. Raw and normalized Zscores are given in the Additional File 1. Zscores varied from the level of noise to a maximum of >20fold the level of noise.
Dimensionality of Isotopomer Data
To determine the dimensionality of the isotopomer data, SVD was performed on ^{13}C fragments derived from uniformly sampled flux distribution sets for several glucose labels. The results are summarized in Figure 6b. Globally, the choice of glucose labels affects the dimensionality of the resulting isotopomer data set. At the 1% (0.01) threshold, the label with the highest dimensionality was hypothetical [1,2,5^{13}C] glucose with 73 dimensions. The three labels for which experimental data was measured in the subsequent section, [1^{13}C] glucose, [6^{13}C] glucose, and 20% [U^{13}C] glucose, had dimensions 53, 53 and 35, respectively, at this cutoff. These values are all significantly lower than the best label, and, in particular, the uniform labeled experiment only produces half of the dimensionality as the optimal experiment. This result is significant. While 139 dimensions (the number of undetermined dimensions for the model used) are required to specify a unique flux vector, the dimensionality of the ^{13}C data for each label is significantly lower. The best labeling experiment specifies just over half (73/139 = 0.52) the degrees of freedom required, and 20% [U^{13}C] glucose only specifies about one fourth of the possible degrees of freedom (0.26). It is worth noting that SVD is a linear operation used to approximate properties of a nonlinear system and the true degrees of freedom may be even lower than reported. SVD serves as a useful upper bound on the dimensionality of data for nonlinear systems, but the difference between SVD dimensionality and true dimensionality may grow to be unacceptable for large systems. For the system studied here, SVD was found to be of practical use.
Experimental Validation
where measured _{ i } is the measured fractional enrichment of fragment i,fragment _{ i }(v) is the computed fractional enrichment of fragment i as a function of the flux distribution v, and α = 0:014 is the standard deviation of the fragments as calculated from experimental replicates.
where c_{ i } = (0, 0, ...0, 1, 0...0) is a vector of all zeros with a 1 in position i, and Error_{ max } was set based on the confidence value. Because different data sets provide different levels of consistency, Error _{ max } was chosen to be 30 units greater than the minimum error found.
At a reaction confidence of 1 mmol · gDW^{1}·h^{1} (a relatively nonstringent cutoff), 85 reactions are specified simply from uptake rate data without any ^{13}C data. Performing the least informative ^{13}C experiment, using 20% [U^{13}C] substrate, yields 105 reactions that meet the confidence criterion, whereas the combination of all three ^{13}C experiments yields 125 reactions that meet the criterion. In other words, performing all three experiments will increase the number of elucidated reactions by 40 reactions or about 50%. As the model used contains 278 carbontracking reactions and reaction groups, the increase in knowledge at 1 mmol · gDW^{1}·h^{1} confidence from 85 to 125 reactions from using ^{13}C data indicates that a large gap in the knowledge remains. It seems apparent that other methods must be developed to obtain flux information at the genome scale from single experiments, as would ultimately be desirable. However, as noted in the above section, the majority of reactions that are elucidated by ^{13}Clabeled glucose experiments lie in central metabolic pathways, which tend to be both of high interest and notspecifiable by constraints alone.
Conclusions
We introduce a new framework for calculating the uncertainty inherent to ^{13}C experiments using Monte Carlo Sampling. This allows us to predict the success of experiments before performing them. The method used here 1) does not require experimental identification of the 'real' flux state a priori [10] and 2) reports scores for the resolution capability for each reaction as opposed to Boolean identification calls [11]. This framework reveals several key findings:

The choice of input label is important, as different labels perform better than others. In particular, the commonlyused 20% mixture of uniform label + 80% natural label was shown computationally and experimentally to resolve significantly fewer reaction fluxes than either [1^{13}C] glucose or [6^{13}C] glucose. Thus, the amount of information likely to be obtained by a ^{13}C experiment can be predicted in a reactionspecific manner before having to carry out an experiment.

There is no universally best label. The best label depends on the experimental objective. Certain reactions are more precisely measured with some labels than others, and no label is best at elucidating all reactions. Certain hypothetical ^{13}C labels of glucose, for example [1,2,3^{13}C] glucose and [1,2,5^{13}C] glucose, are predicted to perform better than commercially available single labels for many reactions.

The ^{13}C data dimensionality is less than anticipated. Whereas each ^{13}C experiment can measure 186 pieces of information at a time, there is a high degree of interdependence. We measured the true data dimensionality to be between 35 and 50 dimensions for commercially available labels and as high as 73 for exotic labels. This high data redundancy can partially explain why ^{13}C experiments yield many reaction rates with high uncertainties.
This study suggests limitations of steadystate ^{13}C analysis using solely amino acids due to the lower than expected dimensionality of the isotopomer data. However, steadystate ^{13}C analysis is clearly still useful elucidating reaction fluxes in E. coli metabolism. Notably, as the study was conducted using only proteinderived amino acids, it would be of immediate interest to determine the additional benefit of measuring other classes of labeled metabolites, as well as the benefit of more recently developed experimental techniques such as multisubstrate [10] and dynamic flux labeling [28, 29] experiments. Monte Carlo methods are generic and thus are well suited to be adapted to the experimental setup of interest. Additionally, Monte Carlo methods are amenable to biasing the sampling space based on known data to improve results; however, this is expected to incur a corresponding cost in convergence time. Possible sources of error in the current method include unwanted bias in the sampled flux space, possible inadequacy of the Zscore as a statistical metric over more sophisticated tests such as the KolmogorovSmirnov test, and inadequacy of the distribution median as a threshold for use in hilo hypotheses. Conducting Monte Carlo isotopomer analysis on new systems will become more accessible with the availability of the open source COBRA Toolbox v2.0 for MATLAB, which includes the algorithms presented here [30].
Methods
Isotopomer Network Description
The isotopomer network was derived from the iMC1010 E. coli reconstruction [19]. The content is reported in the Additional File 1. There are a total of 313 irreversible reactions including 278 that track carbon. All carbon tracking reactions are broken into elementary forward and reverse reactions.
First, a central metabolic isotopomer model was generated that includes a total of 85 reactions, including a biomass production reaction, which drains the precursor metabolites used to make biomass, and 14 system boundary exchange fluxes (for glucose, oxygen, phosphate, NO_{2}, NO_{3}, acetate, CO_{2}, ethanol, formate, fumarate, glycerol, Dlactate, pyruvate, and succinate). The biomass composition is based on one that was reported previously [16, 31] and used in the biosynthetic isotopomer model (see details below), but where the biomass components are replaced by the amount of ATP, NADH, NADPH, and central metabolic precursors needed to synthesize the biomass components (Additional File 1). The remaining 70 reactions participate in glycolysis, TCA cycle, pentose phosphate pathway, oxidative phosphorylation, pyruvate metabolism, and anaplerotic metabolism. The central metabolic isotopomer model includes linear mass balance equations for 67 metabolites. Carbon atoms are tracked through 46 metabolites in the core metabolic network. Changes to the central metabolic reactions include assigning fumarate reductase to utilize menaquinone and demethylmenaquinone rather than ubiquinone and adding a phosphate transport reaction coupled to proton symport.
Aside from the central metabolic reactions contained in the central isotopomer model, the biosynthetic model also includes a number of other catabolic and anabolic reactions. The fluxes were calculated with an additional constraint that flux through formyltetrahydrofolate deformylase (which removes the C1 unit from 10formyltetrahydrofolate) was less than or equal to the measured formate secretion flux. When higher flux through this reaction was allowed the minimum error improved by only 0.3%, but the flux through this reaction was high (around half the glucose uptake rate). To adjust this aberrant behavior, the optimal flux distributions and confidence intervals were calculated with this additional constraint on the formyltetrahydrofolate deformylase flux. The resulting biosynthetic isotopomer model includes 189 metabolites (126 of which have tracked carbon atoms), 313 irreversible metabolic reactions (63 of which are reversible and involve tracked carbon atoms), and 8,612 isotopomer variables (which is equal to the number of nonlinear isotopomer mass balance constraints). The model also includes a biomass reaction and 19 system boundary exchange reactions. The biomass reaction was altered to include amino acids, nucleotides, cofactors, and macromolecules rather than their precursor metabolites. In addition, the biosynthetic model balances intracellular protons as well as water molecules similar to iJR904 [16].
Monte Carlo Sampling
With traditional MCMC, a point is selected within the space which is then iteratively moved around. At each step, a random direction is chosen and the next point is chosen uniformly along this line. The set of points that this algorithm visits will converge to a uniformly distributed set. Two modifications were made: 1) Artificial Centering [32]  Because these biological spaces tend to be elongated in one direction, it is often beneficial to choose directions along the "long" direction rather than uniformly. This can be done by choosing the direction based on previously visited points. At each step, the direction is chosen by drawing a vector from the center of the previous points to one of the previous points chosen at random. 2) In place sampling  Instead of moving just one point throughout the space, many points are moved simultaneously. In this way, no "history" is kept, only the updated position of all the points. This method is described in greater detail in other literature [33, 34].
Computing the Isotopomer Distribution
Each flux distribution and glucose input results in a unique isotopomer distribution. The cumomer method [35] and the elementary metabolite unit (EMU) method [36] were implemented in Matlab and utilized in calculating isotopomer distributions. These methods involve solving several linear systems of equations to compute different groups of isotopomers. For numerical reasons, a routine is introduced which checks whether all parts of the network are still connected at every step. Disconnected components can occur when fluxes to and from the component are zero, making it impossible to compute a unique isotopomer distribution within this subnetwork as many isotopomers satisfy the balance equations. By removing these components first, the other metabolites can be solved in a numerically stable fashion. The resulting isotopomers for the amino acids for each flux distribution are transformed to a mass distribution. This way, each experiment is abstracted to a (number of distributions) × (number of fragments) matrix.
Sample Preparation and ^{13} C Measurement
Culture labeling
Prior to labeling, single colonies of E. coli K12 MG1655 were selected from stock plates and inoculated directly into 250 ml M9 medium in 500 Erlenmeyer flasks aerated by stirring at 1000 rpm. Cells were grown overnight, harvested, washed twice with water and used to inoculate 50 ml flasks containing 25 ml medium with 2 g/L 13Clabeled Dglucose, with initial OD600 0.0050.01. Glucose was supplied as either 100% [1^{13}C]labeled, 100% [6^{13}C]labeled, or a mixture of 20% uniformly [U^{13}C]labeled with 80% natural glucose (which is randomly 1% 13C). Cells were grown to midlog phase, corresponding to OD600 of 0.6. 3 ml of each culture was harvested by centrifugation at 4°C. The media was aspirated and analyzed with HPLC to determine the remaining glucose concentration. Cell pellets were placed at 80°C prior to further analysis.
Derivatization and GCMS analysis
Cells were resuspended in 0.1 ml 6 M HCl and transferred to glass vials. Protein was digested into amino acids under a nitrogen atmosphere for 18 hr at 105°C in an Eldex H/D Work Station. Digested samples were dried to remove residual HCl, resuspended with 75 μl each of tetrahydrofuran and NtertbutyldimethylsilylNmethyltrifluoroacetamide (Aldrich), and incubated for 1 hr at 80°C to derivatize amino acids. Samples were filtered through 0.2 μm PVDF filters and injected into a Shimadzu QP2010 Plus GCMS (0.5 μl with 1:50 split ratio). GC injection temperature was 250°C and the GC oven temperature was initially 130°C for 4 min, rising to 230°C at 4°C/min and to 280°C at 20°C/min with a final hold at this temperature for 2 min. GC flow rate with helium carrier gas was 50 cm/s. The GC column used was a 15 m × 0.25 mm × 0.25 m SHRXI5ms (Shimadzu). GCMS interface temperature was 300 degrees with 70 eV ionization voltage. The mass spectrometer was set to scan an m/z range of 50 to 600.
Processing of GCMS data
Mass data were retrieved from the GCMS for fragments of 14 derivatized amino acids: cysteine and tryptophan were degraded during amino acid hydrolysis; asparagine and glutamine were converted respectively to aspartate and glutamate; arginine was not stable to the derivatization procedure. For each fragment, these data comprised mass intensities for the base isotopomer (without any heavy isotopes, M+0), and isotopomers with increasing unit mass (up to M+6) relative to that of M+0. These mass distributions were normalized by dividing by the sum of M+0 to M+6, and corrected for naturallyoccurring heavy isotopes of the elements H, N, O, Si, S, and (in moieties from the derivatizing reagent) C, using matrixbased probabilistic methods as described [37, 38] implemented in Microsoft Excel. Data were also corrected for carryover of unlabeled inoculum [37].
Computing Reaction Rates from ^{13} C Data
This reduced the number of variables from v = 335 to α = 139. Optimization was performed with the Tomlab/SNOPT package. This method is an iterative local optimization and is therefore not guaranteed to find the optimal solution. To address the issue of local minima, the procedure was run with many randomly generated starting points and the lowest minimum was taken.
Code and Equipment
The code was written in the MATLAB environment and the COBRA toolbox. Linear Programming was done with the Tomlab/CPLEX package and nonlinear optimization with the TOMLAB/SNOPT interface. The EMU and cumomer method were written in native Matlab but generated in Perl. Computations were performed on a Dell Studio XPS desktops (2.6 Ghz core i7 with 912 GB ram) and a custom Rocks cluster (100 dual Xeon 5500 series nodes).
Declarations
Acknowledgements
This work was supported by National Institutes of Health grants GM068837, GM057089, and GM062791.
Authors’ Affiliations
References
 Heiden MGV, Cantley LC, Thompson CB: Understanding the Warburg Effect: The Metabolic Requirements of Cell Proliferation. Science. 2009, 324 (5930): 10291033. 10.1126/science.1160809.View ArticleGoogle Scholar
 Herrgard MJ, Fong SS, Palsson BO: Identification of genomescale metabolic network models using experimentally measured flux profiles. PLoS Comput Biol. 2006, 2 (7): e7210.1371/journal.pcbi.0020072.View ArticleGoogle Scholar
 Wiback S, Mahadevan R, Palsson B: Using Metabolic Flux Data to Further Constrain the Metabolic Solution Space and Predict Internal Flux Patterns: The Escherichia coli alphaSpectrum. Biotechnology and Bioengineering. 2004, 86 (3): 31731. 10.1002/bit.20011.View ArticleGoogle Scholar
 Mollney M, Wiechert W, Kownatzki D, de Graaf AA: Bidirectional Reaction Steps in Metabolic Networks: IV. Optimal Design of Isotopomer Labeling Experiments. Biotechnology and Bioengineering. 1999, 66 (2): 86103. 10.1002/(SICI)10970290(1999)66:2<86::AIDBIT2>3.0.CO;2A.View ArticleGoogle Scholar
 Sauer U: Metabolic networks in motion: 13Cbased flux analysis. Molecular Systems Biology. 2006Google Scholar
 Wiechert W: 13C metabolic flux analysis. Metab Eng. 2001, 3 (3): 195206. 10.1006/mben.2001.0187.View ArticleGoogle Scholar
 Suthers PF, Burgard AP, Dasika MS, Nowroozi F, Van Dien S, Keasling JD, Maranas CD: Metabolic flux elucidation for largescale models using 13C labeled isotopes. Metab Eng. 2007, 9 (56): 387405. 10.1016/j.ymben.2007.05.005.View ArticleGoogle Scholar
 Zamboni N, Fendt SM, Ruhl M, Sauer U: (13)Cbased metabolic flux analysis. Nat Protoc. 2009, 4 (6): 87892. 10.1038/nprot.2009.58.View ArticleGoogle Scholar
 Schmidt K, Carlsen M, Nielsen J, Villadsen J: Modeling Isotopomer Distributions in Biochemical Reaction Networks Using Isotopomer Mapping Matrices. Biotechnology and Bioengineering. 1997, 55 (6): 83140. 10.1002/(SICI)10970290(19970920)55:6<831::AIDBIT2>3.0.CO;2H.View ArticleGoogle Scholar
 Metallo CM, Walther JL, Stephanopoulos G: Evaluation of 13C isotopic tracers for metabolic flux analysis in mammalian cells. J Biotechnol. 2009, 144 (3): 16774. 10.1016/j.jbiotec.2009.07.010.View ArticleGoogle Scholar
 Chang Y, Suthers PF, Maranas CD: Identification of Optimal Measurement Sets for Complete Flux Elucidation in Metabolic Flux Analysis Experiments. Biotechnology and Bioengineering. 2008, 100 (6): 10391049. 10.1002/bit.21926.View ArticleGoogle Scholar
 Price ND, Reed JL, Palsson B: Genomescale models of microbial cells: evaluating the consequences of constraints. Nat Rev Microbiol. 2004, 2 (11): 886897. 10.1038/nrmicro1023.View ArticleGoogle Scholar
 Reed JL, Famili I, Thiele I, Palsson BO: Towards multidimensional genome annotation. Nat Rev Genet. 2006, 7 (2): 13041. 10.1038/nrg1769.View ArticleGoogle Scholar
 Schellenberger J, Park JO, Conrad TM, Palsson BO: BiGG: a Biochemical Genetic and Genomic knowledgebase of large scale metabolic reconstructions. BMC Bioinformatics. 2010, 11: 21310.1186/1471210511213.View ArticleGoogle Scholar
 Thiele I, Palsson BO: A protocol for generating a highquality genomescale metabolic reconstruction. Nat Protoc. 2010, 5: 93121.View ArticleGoogle Scholar
 Reed J, Vo T, Schilling CH, Palsson B: An expanded genomescale model of Escherichia coli K12 (iJR904 GSM/GPR). Genome Biology. 2003, 4 (9): R54.1R54.12.View ArticleGoogle Scholar
 Feist AM, Palsson BO: The biomass objective function. Curr Opin Microbiol. 2010Google Scholar
 Edwards J, Palsson B: The Escherichia coli MG1655 in silico metabolic genotype: Its definition, characteristics, and capabilities. Proc Natl Acad Sci USA. 2000, 97 (10): 55285533. 10.1073/pnas.97.10.5528.View ArticleGoogle Scholar
 Covert MW, Knight EM, Reed JL, Herrgard MJ, Palsson BO: Integrating highthroughput and computational data elucidates bacterial networks. Nature. 2004, 429 (6987): 926. 10.1038/nature02456.View ArticleGoogle Scholar
 Hua Q, Joyce AR, Fong SS, Palsson BO: Metabolic analysis of adaptive evolution for in silico designed lactateproducing strains. Biotechnol Bioeng. 2006, 95 (5): 9921002. 10.1002/bit.21073.View ArticleGoogle Scholar
 Burgard AP, Pharkya P, Maranas CD: Optknock: a bilevel programming framework for identifying gene knockout strategies for microbial strain optimization. Biotechnol Bioeng. 2003, 84 (6): 64757. 10.1002/bit.10803.View ArticleGoogle Scholar
 Feist AM, Zielinski DC, Orth JD, Schellenberger J, Herrgard MJ, Palsson BO: Modeldriven evaluation of the production potential for growthcoupled products of Escherichia coli. Metab Eng. 2009Google Scholar
 Feist AM, Palsson B: The growing scope of applications of genomescale metabolic reconstructions using Escherichia coli. Nat Biotech. 2008, 26 (6): 659667. 10.1038/nbt1401.View ArticleGoogle Scholar
 Schellenberger J, Palsson BO: Use of randomized sampling for analysis of metabolic networks. J Biol Chem. 2009, 284 (9): 545761.View ArticleGoogle Scholar
 Fischer E, Zamboni N, Sauer U: Highthroughput metabolic flux analysis based on gas chromatographymass spectrometry derived 13C constraints. Anal Biochem. 2004, 325 (2): 30816. 10.1016/j.ab.2003.10.036.View ArticleGoogle Scholar
 Zhao J, Shimizu K: Metabolic flux analysis of Escherichia coli K12 grown on 13Clabeled acetate and glucose using GCMS and powerful flux calculation method. J Biotechnol. 2003, 101 (2): 10117. 10.1016/S01681656(02)003164.View ArticleGoogle Scholar
 Schilling CH, Schuster S, Palsson BO, Heinrich R: Metabolic pathway analysis: basic concepts and scientific applications in the postgenomic era. Biotechnology Progress. 1999, 15 (3): 296303. 10.1021/bp990048k.View ArticleGoogle Scholar
 Antoniewicz MR, Kraynie DF, Laffend LA, GonzalezLergier J, Kelleher JK, Stephanopoulos G: Metabolic flux analysis in a nonstationary system: Fedbatch fermentation of a high yielding strain of E. coli producing 1,3propanediol. Metab Eng. 2007Google Scholar
 Yuan J, Bennett BD, Raobinowitz JD: Kinetic flux profiling for quantitation of cellular metabolic fluxes. Nature Protocols. 2008, 3: 13281340. 10.1038/nprot.2008.131.View ArticleGoogle Scholar
 Schellenberger J, Que R, Fleming RMT, Thiele I, Orth JD, Feist AM, Zielinski DC, Bordbar A, Lewis NE, Rahmanian S, Kang J, Hyduke DR, Palsson B: Quantitative prediction of cellular metabolism with constraintbased models: the COBRA Toolbox v2.0. Nature Protocols. 2011, 6: 12901307. 10.1038/nprot.2011.308.View ArticleGoogle Scholar
 Edwards JS, Palsson BO: Multiple steady states in kinetic models of red cell metabolism. Journal of Theoretical Biology. 2000, 207: 1257. 10.1006/jtbi.2000.2165.View ArticleGoogle Scholar
 Kaufmann DE SR: Direction choice for accelerated convergence in hitandrun sampling. Operations Research. 1998, 46: 8495. 10.1287/opre.46.1.84.View ArticleGoogle Scholar
 Lewis NE, Schramm G, Bordbar A, Schellenberger J, Andersen MP, Cheng JK, Patel N, Yee A, Lewis RA, Eils R, Konig R, Palsson BO: Largescale in silico modeling of metabolic interactions between cell types in the human brain. Nature Biotechnology. 2010, 28: 12791285. 10.1038/nbt.1711.View ArticleGoogle Scholar
 Bordbar A, Lewis NE, Schellenberger J, Palsson BO, Jamshidi N: Insight into human alveolar macrophage and M. tuberculosis interactions via metabolic reconstructions. Molecular Systems Biology. 2010, 6 (422):Google Scholar
 Wiechert W, Mollney M, Isermann N, Wurzel M, de Graaf AA: Bidirectional reaction steps in metabolic networks: III. Explicit solution and analysis of isotopomer labeling systems. Biotechnol Bioeng. 1999, 66 (2): 6985. 10.1002/(SICI)10970290(1999)66:2<69::AIDBIT1>3.0.CO;26.View ArticleGoogle Scholar
 Antoniewicz MR, Kelleher JK, Stephanopoulos G: Elementary metabolite units (EMU): a novel framework for modeling isotopic distributions. Metab Eng. 2007, 9: 6886. 10.1016/j.ymben.2006.09.001.View ArticleGoogle Scholar
 Nanchen A, Fuhrer T, Sauer U: Determination of metabolic flux ratios from 13Cexperiments and gas chromatographymass spectrometry data: protocol and principles. Methods Mol Biol. 2007, 358: 17797. 10.1007/9781597452441_11.View ArticleGoogle Scholar
 van Winden WA, Wittmann C, Heinzle E, Heijnen JJ: Correcting mass isotopomer distributions for naturally occurring isotopes. Biotechnol Bioeng. 2002, 80 (4): 4779. 10.1002/bit.10393.View ArticleGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.