Browse Subject Areas

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Physiologically Shrinking the Solution Space of a Saccharomyces cerevisiae Genome-Scale Model Suggests the Role of the Metabolic Network in Shaping Gene Expression Noise

Physiologically Shrinking the Solution Space of a Saccharomyces cerevisiae Genome-Scale Model Suggests the Role of the Metabolic Network in Shaping Gene Expression Noise

  • Baofang Chi, 
  • Shiheng Tao, 
  • Yanlin Liu


Sampling the solution space of genome-scale models is generally conducted to determine the feasible region for metabolic flux distribution. Because the region for actual metabolic states resides only in a small fraction of the entire space, it is necessary to shrink the solution space to improve the predictive power of a model. A common strategy is to constrain models by integrating extra datasets such as high-throughput datasets and C13-labeled flux datasets. However, studies refining these approaches by performing a meta-analysis of massive experimental metabolic flux measurements, which are closely linked to cellular phenotypes, are limited. In the present study, experimentally identified metabolic flux data from 96 published reports were systematically reviewed. Several strong associations among metabolic flux phenotypes were observed. These phenotype-phenotype associations at the flux level were quantified and integrated into a Saccharomyces cerevisiae genome-scale model as extra physiological constraints. By sampling the shrunken solution space of the model, the metabolic flux fluctuation level, which is an intrinsic trait of metabolic reactions determined by the network, was estimated and utilized to explore its relationship to gene expression noise. Although no correlation was observed in all enzyme-coding genes, a relationship between metabolic flux fluctuation and expression noise of genes associated with enzyme-dosage sensitive reactions was detected, suggesting that the metabolic network plays a role in shaping gene expression noise. Such correlation was mainly attributed to the genes corresponding to non-essential reactions, rather than essential ones. This was at least partially, due to regulations underlying the flux phenotype-phenotype associations. Altogether, this study proposes a new approach in shrinking the solution space of a genome-scale model, of which sampling provides new insights into gene expression noise.


As a powerful tool for biological interpretation and discovery, over 100 genome-scale metabolic networks for more than 35 organisms have been reconstructed (see These models have been widely used in molecular evolution studies [13], genome annotation [4, 5], metabolic engineering [69], and cellular phenotype predictions [1012]. The broad applications of genome-scale metabolic network reconstructions largely owe its success to a constraint-based modeling strategy. Instead of developing a system of partial differential equations such as that observed in kinetic models, the constraint-based modeling method converts a metabolic network into a stoichiometric matrix [13]. This matrix ensures balance of metabolites flow throughout the network, thus resulting in hundreds of mass balance constraints. These, together with the upper and lower bounds of the metabolic reactions in the network, define a solution space where actual flux distributions reside [14]. However, in reality, the metabolic behavior of a cell is under complex physiological regulations, many of which have not been reflected by the constraints mentioned above. This implies that the region for biologically relevant flux distribution is represented by small fraction of the entire solution space of a constraint-based model (CBM). Therefore, shrinking the predicted solution space is necessary to improve the predictive ability of a CBM.

Diverse constraint-based reconstruction and analysis (COBRA) methods have been developed to probe the solution space [15]. These could be basically classified into two groups, namely, unbiased and biased methods. Unbiased methods, like the Monte Carlo Markov Chain sampling method, are mainly used to characterize all possible flux distributions within the solution space. However, cells do not utilize most of these flux distributions [16]. On the other hand, Flux Balance Analysis (FBA) is the most extensively employed in biased methods. By optimizing an objective function, FBA attempts to capture the region where real flux distributions are most likely to reside [17]. FBA is capable of performing quantitative predictions with relatively high accuracy. During the past decades, the explosive growth of omics data offered a good opportunity to reduce the solution space of a CBM by integrating experimental datasets, including gene expression data [12,18], proteomic data [19], and metabolite concentration data [20]. In particular, algorithm innovations related to transcriptomic data integration have undergone substantial progress in the past decades [21]. Recently, Herrgård’s group [22] systematically evaluated the predictive capabilities of eight classic methods. The results showed that in many conditions, these methods do not outperform simple FBA by using growth maximization and parsimony criteria when making phenotype predictions. Other studies have also described the limitations of the strategy [12, 23]. Therefore, alternative approaches in locating real flux distributions are warranted.

Extensive experimental evidences suggest that cellular phenotypes could be synergistically regulated by global biological regulatory factors under certain rules [2427]. Compared to mRNA levels, protein abundance, and metabolite concentration in cells, metabolic flux levels are more capable of reflecting cellular phenotypes [28]. Therefore, phenotype-phenotype correlations are very likely to be observed at the metabolic flux level. Exploring metabolic flux phenotype correlations would benefit certain bounds refinements of reactions, further contributing to CBM solution space shrinkage. Experimentally identified metabolic flux data on S. cerevisiae has been rapidly accumulating in the past decades. It has laid a solid foundation on flux phenotype correlation analysis; however, this kind of data has not been extensively utilized. In the present study, a meta-analysis of experimental flux data from publications has been performed to shrink the solution space of the yeast CBM.

A refined model benefits biological discoveries. In the present study, we applied the refined solution space of a S. cerevisiae model to explore the influence of the metabolic network on gene expression noise. Gene expression noise is defined as the stochastic fluctuation of a gene’s expression at the protein level between isogenic cells under identical environment [29]. It is very important to understand the regulatory mechanism of gene expression noise because small variations at the noise level could result in a dramatic change at the phenotype level [30]. Previous studies have indicated that gene expression noise is regulated by several factors, including the structure of genes [31, 32], gene essentiality [31], gene expression level [30, 33], transcription/translation rate [3436], chromatin remodeling [31,37], and regulation network [3840]. A recent study indicated that gene expression noise could propagate and cause fluctuations in growth through corresponding metabolic reactions [41]. However, the interaction between gene expression noise and the metabolic network remains elusive.

In the present study, we first determined the correlations among several metabolic flux phenotypes through a systematic review of experimentally measured metabolic flux data. The correlations were quantified and imposed into a S. cerevisiae CBM to refine its solution space. Using the refined model, we examined whether the metabolic network influences expression noise of enzyme-coding genes.


Correlated metabolic flux phenotype quantification

All the metabolic flux data we collected were measured under the conditions that the yeast strains were cultivated aerobically with glucose as the only carbon source (see Materials and Methods; S1 Dataset). Because sugar uptake rate plays a central role in shaping the global metabolic system under a carbon-limited condition [26], the relationship of other flux phenotypes to glucose uptake rates was explored. In an aerobic condition, S. cerevisiae can be mainly in any of two metabolic states, respiration and fermentation. In the former state, glucose is completely oxidized into CO2 through the mitochondrial respiration pathway, whereas in the latter, glucose is predominantly fermented into ethanol. Because of the existence of two metabolic patterns, it is necessary to identify the metabolic state that the S. cerevisiae is in prior to exploring flux-flux phenotype correlations.

Because the respiratory quotient (RQ) is an indicator of metabolic states, RQ values against glucose uptake rates (For convenience, was used for the y-axes) were plotted. Fig 1 shows that when the glucose uptake rate was <4mmol/ (g (DW)·h), the values fluctuate within the range of 0.81–1.06. This finding was suggestive of the dominant role of respiration over fermentation. When the glucose uptake rate was >4 mmol/ (g (DW)·h), the values gradually decreased as glucose uptake rate increased, which was indicative of a metabolic pattern transition from respiration to the fermentation state. This result is consistent with the study conducted by Postmaet al. [42]. Therefore, the glucose uptake rate “4 mmol/ (g (DW)·h)” was defined as a “transition marker” for discriminating metabolic states.

Fig 1. Distribution of RQ values at different glucose uptake rates ( was used for the y-axis).

Next, O2 uptake rate, ethanol production rate, and glycerol secretion rate against glucose uptake rate were plotted, respectively. Results showed that when glucose uptake rate was <4 mmol/ (g (DW)·h), the O2 uptake rates positively correlated with glucose uptake rates (Fig 2(A)). On the other hand, when the glucose uptake rate was >4 mmol/ (g (DW)·h), the O2 uptake rate negatively correlated with the glucose uptake rate(Fig 2(B)). Unlike the O2uptake rate, ethanol production rates were always positively correlated with glucose uptake rates, independent of the metabolic state of yeast (Fig 2(C)). Glycerol secretion rates were correlated with glucose uptake rates (Fig 2(D)). Based on these findings, the correlation between the three flux phenotypes (O2 uptake rate, ethanol production rate, and glycerol secretion rate) and glucose uptake rate was quantified (Table A in S3 File).

Fig 2. Metabolic flux phenotype-phenotype correlations.

Red circles denote experimentally identified data. Blue lines indicate the upper (lower) bound based on regression analysis result. (A) O2 consumption rate versus glucose uptake rate (<4mmol/ (g (DW)·h)). (B) O2 consumption rate versus glucose uptake rate (>4mmol/ (g (DW)·h). (C) Ethanol production rate versus glucose uptake rate; (D) Glycerol secretion rate versus glucose uptake rate.

Prediction of the upper and lower bounds of the growth rate under aerobic conditions

With the refined bounds for O2 uptake rate, ethanol production rate, and glycerol secretion rate, FBA was employed to predict the upper and lower bounds of growth rate (see Materials and Methods). Fig 3 shows that almost all the experimentally identified growth rates were within the predicted space, indicating that our strategy was suitable for the refinement of flux bounds. In addition, when the glucose uptake rate was <4 mmol/(g(DW)·h), the experimental data fitted well with the predicted upper bound (R2 = 0.83), indicating the dominant role of the respiration pathway in supplying energy for growth. When the glucose uptake rate was >4mmol/ (g(DW)·h), the data points dispersed in the space. This finding suggested that once the nutrients in the environment are sufficient to support a high glucose uptake rate, S. cerevisiae would be relaxed from evolutionary pressures for survival. Under this condition, “biomass growth” is just one of the several goals that a cell strives to achieve [43], and growth rate becomes a yeast strain-dependent phenotype.

Fig 3. Comparison of predicted lower (upper) bound for growth rate and experimentally identified growth rates.

The two blue lines indicate the predicted upper and lower bound for growth rate. The red points denote experimentally identified growth rates. Inset: When the glucose uptake rate was <4 mmol/(g(DW)·h), experimentally identified growth rates fitted well with the predicted upper bound for growth rate.

Metabolic flux moderately correlates with mRNA level and protein abundance

Several studies have revealed a close relationship between metabolic flux and gene expression for various pathways in S. cerevisiae [24, 44, 45], but only a few studies directly identify this relationship [46]. In the present study, the average metabolic flux (AF) level of each metabolic reaction was estimated and compared to the corresponding gene expression level. Table 1 showed that AF moderately correlated with mRNA level and protein abundance, which was consistent the findings of a previous study [46]. Contrary to the results of Bilu [46], our analysis suggested that the correlation between AF and protein abundance level is stronger than that between AF and mRNA level. This result reflected the fact that compared to mRNA level, protein abundance is more closely linked to metabolic flux level.

Table 1. Spearman rank correlation between metabolic flux level and mRNA level/protein abundance.

Metabolic flux fluctuation moderately correlates with gene expression noise

Gene expression noise describes the cell-to-cell differences in protein level while maintaining homeostasis. In 2006, Newman et al. conducted a systematical research on gene expression noise of S. cerevisiae. They proposed DM (which is short for distance to median) value as a measure of gene-specific noise (see Materials and Methods). The noise is able to propagate and cause fluctuations in growth through corresponding metabolic reactions [41]. To know whether the metabolic network has influence on gene expression noise, we estimated metabolic flux fluctuation (FF) level of each reaction. FF could be regarded as reaction's intrinsic trait given by the metabolic network (see Materials and Methods), and was compared with corresponding DM value. Results showed no significant correlation between FF and DM in all enzyme-coding genes (data not shown). A possible reason is that for many reactions, the flux levels are not controlled by enzyme concentrations, but by ubiquitous allosteric regulatory mechanisms [23]. Thus, we next examined the genes that were mostly likely to be related to enzyme-dosage sensitive reactions (DSRs). According to Newman et al., protein abundance levels are measured in both rich (YEPD) and minimal media (SD) [24]. Between the two conditions, several genes were differentially expressed, whereas others remained stable. It is reasonable to assume that the reactions that were catalyzed by enzymes whose level were divergently expressed between the two conditions were more likely to be DSRs. Therefore, the ratio of protein abundance obtained from SD to those obtained from YEPD was used for DSR gene retrieval. Only the genes whose expression levels follow the expression,|Log2(SD/YEPD)| > 1,were considered as DSR genes. Finally, a total of 41 genes were obtained and subjected to correlation analysis. In the DSR genes, FF moderately correlated with DM (rho = 0.27, p = 0.0480; Fig 4(A)). To ensure the robustness of this result, three different cutoffs (log2(SD/YEPD)>0.5,0.7,1.0 respectively) were used to retrieve DSR genes for correlation calculations. The correlations between FF and DM were always positive(rho values ranging from 0.26 to 0.58; p<0.05). The more stringent the cutoff, the stronger the correlation. These findings ascertained that for DSR genes, the level of gene expression noise was subject to the tolerance of the metabolic network to flux fluctuations.

Fig 4. Correlations between flux fluctuation and gene expression noise.

(A) Correlation between FF and DM for DSR genes. (B) Correlation between FF and DM for NER genes. (C) Correlation between FF and DM for ER genes.

Reaction essentiality is related to DM and FF coupling

Gene expression noise undergoes selection pressures when it is beneficial or detrimental to the survival of an organism [3032]. However, that selection pressure for traits that are not vital to cell growth may be relaxed. To determine potential influences of reaction essentiality on FF and DM correlation, the 41 DSR genes were subdivided into “essential reaction” and “non-essential reaction” groups. Among the 41 genes, 20 were associated with essential reactions, whereas 21 were associated with non-essential reactions. Correlation analysis indicated that for the genes corresponding to non-essential reactions, FF strongly correlated with DM (rho = 0.60, p = 0.0050; Fig 4(B)). However, we did not observe a significant correlation in the genes that corresponded to essential reactions (Fig 4(C)).

The DM and FF values between the essential and non-essential reaction groups were also compared. While no obvious differences were observed in the average DM values between the two groups, the FF values of non-essential reactions were generally higher than those of essential reactions (Fig 5). A simulation test was also performed to ensure the robustness of this result (see Materials and Methods), which showed that the simulated difference values between two groups were normally distributed around 0, with a standard deviation σ = 0.26. The observed difference value was at least four standard deviations away from the simulated mean difference. Therefore, the observation that FF values of non-essential reactions were higher than those of essential ones occurred by chance was highly unlikely (Fig 6).

Fig 5. The FF values of the non-essential reaction (NER) group were generally higher than those of the essential reaction (ER) group.

Fig 6. The observed differences in average FF values between the ER and NER groups are statistically significant.

The x-axis shows the average FF difference between the ER group and the NER group in 10,000 random simulations. The y-axis indicates the number of simulated FF differences in a bin of width 0.1.

To avoid the possibility that all the results obtained in this section were sampling artifacts of a specific part or the entire solution space, FF values were re-calculated by sampling another part of the shrunken space where the glucose uptake rate fluctuates within 20-22mmol/(g (DW)·h). These FF values were also used in the correlation analysis, and the results were the same as those obtained earlier (Fig B in S3 File).

Physiological regulations mitigate FF and DM coupling

In the previous section, we revealed the difference between “non-essential reaction”-associated and “essential reaction”-associated genes in FF and DM coupling. The difference was speculated to be related to the flux phenotype-phenotype correlations that were quantified and imposed into the yeast model as additional constraints. To test this hypothesis, the FF values of DSR genes were re-calculated by sampling the yeast model’s solution space without the constraints. Removal of the constraints did not remarkably affect the strong correlation with the “non-essential reaction”-related genes (rho = 0.59, p = 0.0051; Fig 7(A)). However, for genes related to essential reactions, newly calculated FF values strongly correlated with the DM values (rho = 0.47, p = 0.037; Fig 7(B)). These findings indicated that the global regulations underlying the phenotype-phenotype correlations mitigated the correlation between FF and DM in genes corresponding to essential reactions.

Fig 7. (A) Correlation between FF and DM in NER genes. (B) Correlation between FF and DM in ER genes. The FF values were calculated by sampling the entire yeast CBM solution space without the extra constraints.

Flux constraint strength is not involved in expression noise regulation

In previous section, we have revealed that metabolic network has influence on gene expression noise. To know whether the influence is related to the extra constraint strength, we used FCS to estimate the level of flux constraints posed on each reaction (see Materials and Methods). Fig 8 showed that the majority of FCS values range within 0–0.1 (746/885), indicating the broad effect of the regulatory mechanism underlying the constraints. However, we did not observed a significant correlation between flux constraint strength and gene expression noise (Fig C in S3 File). This implied that the regulatory strength does not play a key role in shaping gene expression noise.

Fig 8. The number of genes under different constraint strengths.

The smaller the FCS value, the greater the constraint strength.


Several studies have identified the limitations of using transcriptomic data in refining CBMs because complex regulations exist during transcription and translation. Because microbial metabolism is operated under certain global principles [2427], we hypothesized that metabolic flux phenotypes are also under synergistical regulations, and thus correlate with each other. By summarizing the metabolic data that have accumulated in the past decades, we have identified several metabolic flux phenotype correlations, which were subsequently and successfully applied to the refinement of solution space. We expected to identify additional metabolic phenotype associations, for which the accumulated metabolic data during the past decades has served as solid foundation.

FF is an intrinsic trait of a metabolic reaction that is determined by the metabolic network where a reaction is located. In the present study, FF was found to correlate with DM for the DSR genes, suggesting that the metabolic network influenced gene expression noise. In particular, DSR genes corresponding to NERs showed a strong correlation between flux fluctuation and gene expression noise, which, however, was not observed in those related to essential reactions. This is, at least partially, due to the underlying flux phenotype-phenotype associations, because after the removal of the extra constraints, FF strongly correlated with DM for the DSR genes related to essential reactions. We thus concluded that the metabolic network is also an important determinant that constrains gene expression noise, especially for genes associated with the reactions that are not essential for organism's survival. However, for reactions that are vital to cell growth, stochastic fluctuations at the flux level led by gene expression noise might be detrimental. Under this condition, the cell adopts additional physiological regulatory mechanisms to precisely, rather than strongly, control gene expression, and thus reduce expression noise. By using this physiological response, the connections between the metabolic network and gene expression noise is mitigated.

In summary, the present study has revealed the associations among several metabolic flux phenotypes through S. cerevisiae physiological data mining. The integration of these flux phenotype-phenotype correlations into a yeast model has resulted in a substantial shrinkage of its solution space. By analyzing the biologically more relevant solution space, we concluded that the metabolic network contributes, sometimes predominantly, to shaping gene expression noise.

Materials and Methods

Data sources

Metabolic flux data.

A comprehensive search of English-language published reports was performed using three search engines (PubMed, Google Scholar, and ScienceDirect). Several publisher archives, including “Applied and Environmental Microbiology”, “Yeast”, “Enzyme and Microbial Technology”, “Applied Microbiology Biotechnology”, and “Microbial Cell Factories”, were also thoroughly surveyed. The search terms included "Saccharomyces cerevisiae" or "S. cerevisiae" or "Yeast," plus “physiology” or "physiological," plus “metabolic” or “metabolism”. After removing duplicates, a total of 160 full-text articles were obtained. The content of these articles were reviewed and 96 of these were accepted (S2 File). The retrieved metabolic data met following criteria:

  1. The unit of measurement for growth rate is “h−1”; The units of measurement for other metabolic flux data is (or could be converted to)“mmol/ (g (DW)·h)”;
  2. The yeast strains used in experiments did not undergo any artificial genetic manipulation at the molecular level such as gene deletion or overexpression;
  3. All the wild-type yeast strains were cultivated aerobically in a steady state; and
  4. Glucose was the only carbon source for yeast, and carbon was the only nutrient limitation of the culture.

After selection according to these criteria, we aggregated and normalized the data to generate a metabolic flux dataset (S1 Dataset), which includes glucose uptake rates, O2 uptake rates, CO2 production rates, respiratory quotient (RQ) values, ethanol production rates, and glycerol secretion rates.

Protein abundance.

We considered three quantitative genome-wide measurements of protein abundance: genome-scale data from the study conducted by Ghaemmaghami et al. [47], and two genome-scale measurements in two conditions as performed by Newman et al. [29].

mRNA level.

We considered three genome-wide measurements of mRNA levels according to Holstege et al. [48], Ingolia et al. [49], and Wang et al. [50].

Gene expression noise.

We considered two yeast genome-wide datasets of gene expression noise level based on the report of Newman et al. [29]. In their work, DM was used as a measure of gene expression noise. DM was defined as the difference of the gene specific noise and the median noise for proteins with the same abundance. In this way, DM represents gene-specific noise independent of the corresponding gene expression level.

For protein abundance, mRNA levels, and gene expression noise, we averaged across data sets (after normalizing each data set by its mean) to minimize experimental noise.

Quantification of metabolic flux phenotype-phenotype correlations and identification of flux boundaries

We examined the relationship between the glucose uptake phenotype and three metabolic flux phenotypes, namely,O2 uptake rate, ethanol production rate, and glycerol secretion rate, respectively.

Because O2 uptake rate and ethanol production rate linearly well correlate with glucose uptake rate (Fig 2), we learned linear models for them. To ensure robustness and avoid publication bias in the linear model, a procedure analogous to leave-one-out cross-validation was performed [51]: The data from each paper were regarded as elementary dataset. All these datasets were aggregated to construct a “reference dataset”. Next, leave-one-elementary dataset-out (LOE) datasets (LDs) were constructed by iteratively omitting each elementary dataset from the reference dataset (If m elementary datasets are available for analysis, then m LOE datasets will be obtained). Linear regression analysis was conducted using the data from each LOE dataset, and m slope values and corresponding R2 values were obtained. The results showed that the m slope values were almost equal to each other, with R2>0.8, indicating the robustness of our modeling strategy.

By performing linear regression analysis, the slope value α (αoxy for oxygen uptake versus glucose uptake curve; and αeth for ethanol production versus glucose uptake curve) was defined. While fixing the α value, two β values, namely, βl and βu, were set through visual inspection to define a space covering > 90% of the following experimental identified points: where vglu represents glucose uptake rate; and lb (ub) represents lower (upper) bound for ethanol (or O2) exchange rate.

Glycerol secretion rate did not correlate well with glucose uptake rate. Therefore, its bounds were defined, somewhat arbitrarily, by modifying the α value as follows: where lbgly (ubgly)represents the lower (upper) bound of the glycerol secretion rate.

All aforementioned α and β values are listed in Table A in S3 File. The newly refined glucose uptake-dependent bounds for O2 consumption rate, ethanol production, rate, and glycerol secretion rate are as follows: (1) (2) (3)

Prediction of the upper and lower bound for growth rate

S. cerevisiae genome-scale metabolic network model iMM904 was used because of its high predictive power among available yeast models [52, 53]. To simulate yeast cultivated in glucose-limited media, fructose and ethanol uptake rates were fixed at 0 because these are potential carbon sources [52]. At each glucose uptake rate, voxy, veth and vgly values were uniformly sampled within the bounds defined by (1), (2), and (3), respectively, and the maximum growth vgro rate that the yeast could achieve was calculated through FBA as follows: (4) subject to: S•v = 0 where vgro indicates the flux of “biomass growth”;S is an M × N stoichiometric matrix (M is the number of metabolites, and N is the number of reactions);v is a vector containing metabolic flux vi of N reactions; and lbi (ubi) represents the lower (upper) bound for vi. When lboxy <0.016 mmol/ (g (DW)·h) at a certain glucose uptake rate, we set lboxy to 0.016 mmol/ (g (DW)·h) [5].

After repeating this process 10,000 times, the biggest (smallest) value in the 10,000 simulated growth rate values was regarded as the maximum (minimum) growth rate vmax (vmin) the yeast was able to achieve under a glucose uptake rate. The vmax (vmin) under glucose uptake rates ranging from 0 to 24 (step size = 0.1) were all computed.

Because the upper (lower) bound for growth rate was likely to be a combination of several linear functions (Fig 3), the predicted vmax (vmin) was divided into several segments according to glucose uptake rates, for each of which linear regressions of the data were performed separately (Fig A in S3 File).The resulting equations were used as upper (lower) bound for the growth rate (All the equations are listed in Table B in S3 File). For simplicity, the bounds were presented as follows: (5)

Estimation of the average flux, flux fluctuation and flux constraint strength

The average flux level (AF), flux fluctuation (FF), and flux constraint strength (FCS) of each metabolic reaction were estimated by sampling S. cerevisiae's solution space. The sampling was restricted to the part of solution space where glucose uptake rate fluctuates within 18-20mmol/ (g (DW)·h) for two considerations: (1) Fluctuations of catabolically active enzymes could lead to growth fluctuations [41]; and (2) Data on mRNA levels to which AF would be compared, and data for gene expression noise to which FF would be compared, were all collected from rich media, under which glucose uptake rate ranges within 0.36–0.40 [54]. The upper (lower) bound for voxy, veth and vgly were set according to (1), (2), and (3), respectively. Facilitated with these additional constraints, The shrunken solution space was fully sampled to generate 500,000 eligible flux distributions.

The sampled flux values of each metabolic reaction were extracted from the 500,000 flux distributions. Given a reaction, the top and bottom 5% simulated flux values were discarded. The remaining sampled flux values were used to calculate AF and FF as follows: where AFi is defined as average flux level of the i-threaction; and FFi is the flux fluctuation of the i-th reaction. vi,max (vi,min)represents the maximum (minimum) flux value among the sampled flux values of the i-th metabolic reaction.

To estimate the constraint strength posed on each reaction, we also sampled yeast model's solution space 500,000 times before imposing the extra constraints. After removing the top (bottom) 5% flux values of each reaction, FCS was defined as follows: where vi,max,ori (vi,min,ori) represents the i-th maximum (minimum) flux value, which was generated from the original solution space. Reactions and metabolites participating in the intracellular loops were not included in the analysis [45]. To exclude potential biases, reactions catalyzed by isozymes were also excluded in the correlation analysis, thus ensuring a one gene-one reaction relationship. Bi-directional reactions were also not involved in AF and FF calculation.

Identification of essential and non-essential reactions

Setting “biomass growth” as the objective function to be optimized, FBA was conducted to predict the growth rate vori of yeast model iMM904 without any extra constraints. We iteratively set reaction flux vi in the yeast model to 0 and employed FBA again to predict the growth rate vgro,i. When vgro,i = 0, this i-th reaction was defined as an “essential reaction.” When vgro,i = vori, this i-th reaction was defined as a “non-essential reaction.”

Simulation test for FF value differences

FF values of the 41 DSR genes were shuffled and randomly divided into two groups, one consisting of 20 values and the remaining values were binned into another one. Difference in average FF values was calculated by subtracting the average FF value in one group by that of another. This process was repeated 100,000 times.

Statistical analysis

All simulations and statistical analyses were performed in MATLAB, which was interfaced with the COBRA toolbox and ‘glpk’ solver. All correlations described in this report are Spearman’s rank correlations.

Supporting Information

S1 Dataset. Normalized metabolic flux data used for analysis.


S1 File. Flowchart of the selection process.


S2 File. The publications used for S. cerevisiae metabolic data extraction.


S3 File. Supplementary figures and tables.

Regression analysis of growth rate against glucose uptake rate in each segment (Figure A).Correlations between metabolic flux and gene expression noise (Figure B).Relationship between metabolic flux constraint (FSC) and gene expression noise (DM) (Figure C).The newly defined bounds (Table A).Regression analysis of growth rate versus glucose uptake rate in each segments. (y = αx +β; y is for growth rate and x is for glucose uptake rate(Table B).



We are grateful to all members of Bioinformatics Centre for many insightful discussions. We also thank Zhimin Xu for his excellent suggestions on the manuscript.

Author Contributions

Conceived and designed the experiments: BC. Performed the experiments: BC. Analyzed the data: BC ST. Contributed reagents/materials/analysis tools: BC ST. Wrote the paper: BC YL.


  1. 1. Ibarra RU, Edwards JS, Palsson BØ (2002) Escherichia coli K-12undergoes adaptive evolution to achieve in silico predicted optimal growth. Nature 420: 186–189. pmid:12432395
  2. 2. Fong SS, Burgard AP, Herring CD, Knight EM, Blattner FR, Maranas CD, et al. (2005)In silico design and adaptive evolution of Escherichia coli for production of lactic acid. Biotechnol Bioeng 91:643–648. pmid:15962337
  3. 3. Pal C, Papp B, Lercher MJ, Csermely P, Oliver SG, Hurst LD, et al. (2006) Chance and necessity in the evolution of minimal metabolic networks. Nature 440: 667–670. pmid:16572170
  4. 4. Reed JL, Patel TR, Chen KH, Joyce AR, Applebee MK, Herring CD, et al. (2006) Systems approach to refining genome annotation. Proc Natl Acad Sci USA 103: 17480–17484. pmid:17088549
  5. 5. Kumar VS, Dasika MS, Maranas CD (2007) Optimization based automated curation of metabolic reconstructions. BMC Bioinformatics 8: 212. pmid:17584497
  6. 6. Park JM, Kim TY, Lee SY (2009) Constraints-based genome-scale metabolic simulation for systems metabolic engineering. Biotechnol Adv 27: 979–988. pmid:19464354
  7. 7. Ng CY, Jung MY, Lee J, Oh MK (2012) Production of 2,3-butanediol in Saccharomyces cerevisiae by in silico aided metabolic engineering. Microb Cell Fact 11: 68. pmid:22640729
  8. 8. Xu P, Ranganathan S, Fowler ZL, Maranas CD, Koffas MA (2011)Genome-scale metabolic network modeling results in minimal interventions that cooperatively force carbon flux towards malonyl-CoA. Metab Eng 13: 578–587. pmid:21763447
  9. 9. Kim J, Reed JL (2010) OptORF: Optimal metabolic and regulatory perturbations for metabolic engineering of microbial strains. BMC Syst Biol 4: 53. pmid:20426856
  10. 10. Fong SS, Palsson BØ (2004) Metabolic gene-deletion strains of Escherichia coli evolve to computationally predicted growth phenotypes. Nat Genet36: 1056–1058. pmid:15448692
  11. 11. Schuetz R, Kuepfer L, Sauer U (2007) Systematic evaluation of objective functions for predicting intracellular fluxes in Escherichia coli. Mol Syst Biol3: 119. pmid:17625511
  12. 12. Rossell S, Huynen MA, Notebaart RA (2013) Inferring Metabolic States in Uncharacterized Environments Using Gene-Expression Measurements. PLoS Comput Biol 9: e1002988. pmid:23555222
  13. 13. Bordbar A, Monk JM, King Z, Palsson BO (2014) Constraint-based models predict metabolic and associated cellular functions. Nat Rev Genet 15: 107–120. pmid:24430943
  14. 14. Edwards JS, Covert M, Palsson BO (2002) Metabolic modelling of microbes: The flux-balance approach. Environ Microbiol 4: 133–133. pmid:12000313
  15. 15. Lewis NE, Nagarajan H, Palsson BO (2012) Constraining the metabolic genotype-phenotype relationship using a phylogeny of in silico methods. Nat Rev Microbiol 10: 291–305. pmid:22367118
  16. 16. Schellenberger J, Palsson BØ (2008) Use of randomized sampling for analysis of metabolic networks. J Biol Chem 284: 5457–5461. pmid:18940807
  17. 17. Orth JD, Thiele I, Palsson BØ (2010) What is flux balance analysis? Nat Biotechnol 28: 245–248. pmid:20212490
  18. 18. Zur H, Ruppin E, Shlomi T (2010) iMAT: an integrative metabolic analysis tool. Bioinformatics 26: 3140–3142. pmid:21081510
  19. 19. Yizhak K, Benyamini T, Liebermeister W, Ruppin E, Shlomi T (2010) Integrating quantitative proteomics and metabolomics with a genome-scale metabolic network model. Bioinformatics 26: 255–260.
  20. 20. Henry CS, Broadbelt LJ, Hatzimanikatis V (2007) Thermodynamics-based metabolic flux analysis. Biophys J. 92: 1792–1805. pmid:17172310
  21. 21. Blazier A, Papin J (2012) Integration of expression data in genome-scale metabolic network reconstructions. Front Physiol 3: 1–7.
  22. 22. Machado D, Herrgård M (2014) Systematic Evaluation of Methods for Integration of Transcriptomic Data into Constraint-Based Models of Metabolism. PLoS Comput Biol 10(4): e1003580. pmid:24762745
  23. 23. Chubukov V, Uhr M, Chat LL, Kleijn RJ, Jules M, Link H, et al. (2013) Transcriptional regulation is insufficient to explain substrate-induced flux changes in Bacillus subtilis. Mol Syst Biol 9:709. pmid:24281055
  24. 24. Bordel S, Agren R, Nielsen J (2010) Sampling the Solution Space in Genome-Scale Metabolic Networks Reveals Transcriptional Regulation in Key Enzymes. PLoS Comput Biol 6(7): e1000859. pmid:20657658
  25. 25. Bordel S (2013) Experimental evidence suggests the existence of evolutionary global operation principles governing microbial metabolism. Sci Rep. 3, 3017; pmid:24145501
  26. 26. Huberts DH, Niebel B, Heinemann M (2012) A flux-sensing mechanism could regulate the switch between respiration and fermentation. FEMS Yeast Res 12(2012) 118–128.
  27. 27. Ostergaard S, Walløe KO, Gomes CSG, Olsson L, Nielsen J (2001) The impact of GAL6, GAL80 and MIG1 on glucose control of the GAL system in Saccharomyces cerevisiae. FEMS Yeast Res 1: 47–55. pmid:12702462
  28. 28. Moxley JF, Jewett MC, Antoniewicz MR, Villas-Boas SG, Hal Alper H, Wheeler RT, et al. (2009) Linking high-resolution metabolic flux phenotypes and transcriptional regulation in yeast modulated by the global regulator Gcn4p. Proc Natl Acad Sci USA 106(16): 6477–6482. pmid:19346491
  29. 29. Newman JR, Ghaemmaghami S, Ihmels J, Breslow DK, Noble M, DeRisi JL, et al. (2006) Single-cell proteomic analysis of S. cerevisiae reveals the architecture of biological noise. Nature 441:840–846. pmid:16699522
  30. 30. Dong D, Shao X, Deng N, Zhang Z (2011) Gene expression variation are predictive for stochastic noise. Nucleic Acids Res 39(2): 403–413. pmid:20860999
  31. 31. Lehner B (2010) Conflict between Noise and Plasticity in Yeast. PLoS Genet 6(11): e1001185. pmid:21079670
  32. 32. Shalem O, Carey L, Zeevi D, Sharon E, Keren L, Weinberger A, et al. (2013) Measurements of the Impact of 3’ End Sequences on Gene Expression Reveal Wide Range and Sequence Dependent Effects. PLoS Comput Biol 9(3): e1002934. pmid:23505350
  33. 33. Bar-Even A, Paulsson J, Maheshri N, Carmi C, O’Shea E, Pilpel Y, et al. (2006) Noise in protein expression scales with natural protein abundance. Nat Genet 38(6):636–643. pmid:16715097
  34. 34. Salari R, Wojtowicz D, Zheng J, Levens D, Pilpel Y, Przytycka TM (2012) Teasing Apart Translational and Transcriptional Components of Stochastic Variations in Eukaryotic Gene Expression. PLoS Comput Biol 8(8): e1002644. pmid:22956896
  35. 35. Blake WJ, Kaern M, Cantor CR, Collins JJ (2003) Noise in eukaryotic gene expression. Nature 422: 633–637. pmid:12687005
  36. 36. Ozbudak EM, Thattai M, Kurtser I, Grossman AD, van Oudenaarden A (2002) Regulation of noise in the expression of a single gene. Nat Genet, 31, 69–73. pmid:11967532
  37. 37. Raser JM, O’Shea EK (2004) Control of stochasticity in eukaryotic gene expression. Science 304: 1811–1814. pmid:15166317
  38. 38. Chalancon G, Ravarani CNJ, Balaji S, Martinez-Arias A, Aravind L, Jothi R, et al. (2012) Interplay between gene expression noise and regulatory network architecture. Trends Genet 28(5): 221–232. pmid:22365642
  39. 39. Murphy KF, Adams RM, Wang X, Balázsi G, Collins JJ (2010)Tuning and controlling gene expression noise in synthetic gene networks. Nucleic Acid Res 38(8): 2712–2726. pmid:20211838
  40. 40. Shahrezaei V, Ollivier JF and Swain PS (2008) Colored extrinsic fluctuations and stochastic gene Expression. Mol Syst Biol 4:196. pmid:18463620
  41. 41. Kiviet DJ, Nghe P, Walker N, Boulineau S, Sunderlikova V, Tans SJ (2014) Stochasticity of metabolism and growth at the single-cell level. Nature 514: 376–379. pmid:25186725
  42. 42. Postma E, Verduyn C, Scheffers WA, Vandijken J (1989) Enzymic analysis of the crabtree effect in glucose-limited chemostat cultures of Saccharomyces cerevisiae. Appl Environ Microbiol 55(2): 468–477. pmid:2566299
  43. 43. Schuetz R, Zamboni N, Zampieri M, Heinemann M, Sauer U (2012) Multidimensional optimality of microbial metabolism. Science 336: 601–604. pmid:22556256
  44. 44. Daran-Lapujade P, Jansen ML, Daran JM, van Gulik W, de Winde JH, Pronk JT (2004) Role o f transcription a l regulation in con trolling fluxes in central carbon metabolism of Saccharomyces cerevisiae. A chemostat culture study. J Biol Chem 279: 9125–9138. pmid:14630934
  45. 45. Mo M, Palsson BO, Herrgard M (2009) Connecting extracellular metabolomic measurements to intracellular flux states in yeast. BMC Syst Biol 3: 37. pmid:19321003
  46. 46. Bilu Y, Shlomi T, Barkai N, Ruppin E (2006) Conservation of expression and sequence of metabolic genes is reflected by activity across metabolic states. PLoS Comput Biol 2(8): e106. pmid:16933982
  47. 47. Ghaemmaghami S, Huh WK, Bower K, Howson RW, Belle A, Dephoure N, et al. (2003) Global analysis of protein expression in yeast. Nature 425: 737–741. pmid:14562106
  48. 48. Holstege FC, Jennings EG, Wyrick JJ, Lee TI, Hengartner CJ, Green MR, et al. (1998) Dissecting the regulatory circuitry of a eukaryotic genome. Cell 95: 717–728. pmid:9845373
  49. 49. Ingolia NT, Ghaemmaghami S, Newman JR, Weissman JS (2009) Genome-wide analysis in vivo of translation with nucleotide resolution using ribosome profiling. Science 324: 218–223. pmid:19213877
  50. 50. Wang Y, Liu CL, Storey JD, Tibshirani RJ, Herschlag D, Brown PO (2002) Precision and functional specificity in mRNA decay. Proc Natl Acad Sci USA 99: 5860–5865. pmid:11972065
  51. 51. Kotelnikova E, Shkrob MA, Pyatnitskiy MA, Ferlini A, Daraselia N (2012) Novel approach to meta-analysis of microarray datasets reveals muscle remodeling-related drug targets and biomarkers in duchenne muscular dystrophy. PLoS Comput Biol 8(2): e1002365. pmid:22319435
  52. 52. Chandrasekaran S, Price ND (2013) Metabolic Constraint-Based Refinement of Transcriptional Regulatory Networks. PLoS Comput Biol 9(12): e1003370. pmid:24348226
  53. 53. Zomorrodi AR, Maranas CD (2010) Improving the iMM904 S. cerevisiae metabolic model using essentiality and synthetic lethality data. BMC Syst Biol 4:178. pmid:21190580
  54. 54. Heyland J, Fu J, Blank LM (2009) Correlation between TCA cycle flux and glucose uptake rate during respiro-fermentative growth of Saccharomyces cerevisiae. Microbiol155: 3827–3837.