Broad metabolic sensitivity profiling of a prototrophic yeast deletion collection
© VanderSluis et al.; licensee BioMed Central Ltd. 2014
Received: 7 December 2013
Accepted: 10 April 2014
Published: 10 April 2014
Genome-wide sensitivity screens in yeast have been immensely popular following the construction of a collection of deletion mutants of non-essential genes. However, the auxotrophic markers in this collection preclude experiments on minimal growth medium, one of the most informative metabolic environments. Here we present quantitative growth analysis for mutants in all 4,772 non-essential genes from our prototrophic deletion collection across a large set of metabolic conditions.
The complete collection was grown in environments consisting of one of four possible carbon sources paired with one of seven nitrogen sources, for a total of 28 different well-defined metabolic environments. The relative contributions to mutants' fitness of each carbon and nitrogen source were determined using multivariate statistical methods. The mutant profiling recovered known and novel genes specific to the processing of nutrients and accurately predicted functional relationships, especially for metabolic functions. A benchmark of genome-scale metabolic network modeling is also given to demonstrate the level of agreement between current in silico predictions and hitherto unavailable experimental data.
These data address a fundamental deficiency in our understanding of the model eukaryote Saccharomyces cerevisiae and its response to the most basic of environments. While choice of carbon source has the greatest impact on cell growth, specific effects due to nitrogen source and interactions between the nutrients are frequent. We demonstrate utility in characterizing genes of unknown function and illustrate how these data can be integrated with other whole-genome screens to interpret similarities between seemingly diverse perturbation types.
Large scale gene deletion screens have become common in Saccharomyces cerevisiae due to efforts in the yeast community to assemble a near complete collection of non-essential single-mutant strains . The subsequent refinement of mating-based high-throughput strain construction techniques such as synthetic genetic array (SGA) analysis  has further driven the creation of customized yeast deletion arrays. While quantitative single mutant fitness assays have been performed , they are generally limited to a single growth medium. A few notable exceptions have begun to explore this space [4–7], but the conditions of interest are often chosen with human therapeutic ends in mind and are limited to known drugs or small molecules of unknown biological effect. A decade and a half after the sequencing of the best-studied eukaryote, a systematic exploration of mutant growth across basic nutrient environments is conspicuously absent. These data would be valuable for metabolic researchers and computational biologists that attempt to model the metabolic network of the cell using methodologies such as flux balance analysis (FBA)  because the defined growth conditions are amenable to modeling.
Yeast strain collections used in previous high-throughput assays (that is, the deletion collection) are auxotrophic , and therefore unable to survive in minimal media unless provided additional nutrients. This requirement reflects the historical use of auxotrophic markers for genetic selection. The resulting requirement for nutrient supplementation precludes systematic testing of the yeast deletion collection on specific combinations of carbon and nitrogen sources because the auxotrophic nutrient supplements can also be used as carbon and nitrogen sources. Previous work has shown not only that nutrient supplementation can have different physiological consequences from genetic complementation  but also that auxotrophies can alter the expression of many other genes .
Summary of conditions and hits called in each condition
Results and discussion
Prototrophic deletion set construction and profiling
Briefly, a MATα strain carrying the SGA marker [11, 12] was crossed to the MAT a yeast deletion set , selected for diploids, and sporulated. Prototrophic haploids were selected using the SGA approach . The final genotype of these 4,772 strains is MAT a yfgΔ0::KanMX can1Δ::STE2pr-SpHIS5 his3Δ1 lyp1Δ0. These strains were then pinned out onto plates containing one of four different carbon sources along with one of seven nitrogen sources. All 28 carbon:nitrogen combinations were included to produce a broad set of well-defined metabolic conditions. The plates were imaged in time course in order to estimate growth rates from measurements of colony size (Figure 1; see Materials and methods for details; Additional file 1).
Yeast wild-type growth suggests carbon/nitrogen interactions
Fitness determination of deletion mutants over the media conditions
In an effort to identify mutant growth defects specific to particular conditions, we derived a model designed to score growth rate for each deletion strain in a given condition relative to its growth under a reference condition (glucose:ammonium). First, the growth rate data (Additional file 1) were normalized for each experimental condition with respect to the glucose:ammonium reference (see Materials and methods). This controlled for the growth rate differences observable in wild-type cells across the different conditions (Figure 2; Additional file 2) and enabled us to focus on more subtle effects due only to the genetic perturbation. A modified z-score was then calculated for each mutant strain (see Materials and methods; Additional file 3). This measure is negative if the strain grew slower in the test condition than would be expected due to the nutrient environment alone, and positive if the strain grew faster than expected. The distribution of growth rates in the 701 wild-type replicates was used to assess the statistical significance of mutant effects in each condition and estimate a false discovery rate (FDR) for any gene-environment interactions (see Materials and methods; Additional file 4). Table 1 shows the number of deletions that grew slower or faster than expected at an FDR threshold of 20% (see Additional file 3 for a complete list of z-scores). While the large number of wild-type replicates allowed for confidence in the small differences in reference strain growth between various nitrogen sources when paired with glycerol, the mutant data on glycerol proved to be too noisy due to extremely slow growth to call mutant effects. Therefore, no growth rate (z-score) data are presented for mutant strains on glycerol.
Observations in galactose concur with previous auxotrophic studies
Another possible explanation for differences between our galactose results and those from the Dudley et al. study is the absence of antimycin A in our media. Antimycin A inhibits energy production from respiratory pathways and forces the strains to ferment galactose. In our experiments, yeast had access to oxygen and could perform both respiration and fermentation with galactose as a carbon source, which is the natural metabolism of galactose by S. cerevisiae.
Liquid culture validation of mutant fitness measurements
We independently validated our single mutant fitness measurements by measuring the growth rate of 40 mutants in a liquid growth assay performed across 20 of the experimental conditions (excluding ribose:arginine and all glycerol pairings; see Materials and methods). The overall correlation between wild-type strain growth rates from these two different approaches was 0.65 (P < 0.003; Pearson), suggesting general agreement between growth rates determined on solid and liquid media. We then adjusted the liquid growth scores, controlling for the wild-type rate in the given condition and the relevant mutant rate in glucose:ammonium so they would reflect condition-specific effects, similar to our modified z-score derived from the agar experiment. The Spearman rank correlation between the adjusted liquid growth score and our agar z-score (for 40 mutant strains × 19 conditions) was 0.34 (P < 2.2 × 10−16). Further excluding glucose conditions (which are generally sparser in the z-score data as a consequence of our use of glucose:ammonium as a reference) increases this correlation to 0.38. Thus, we conclude that there is reasonable agreement between the high-throughput measures and a lower-throughput liquid growth assay, including for condition-specific effects.
Number of environmental sensitivities is correlated with single mutant fitness and genetic interaction degree
We compared our growth measurements with other quantitative phenotypes measured on the auxotrophic deletion collection. For example, genetic interaction mapping efforts have measured the single mutant fitness of all deletion strains from the auxotrophic background on synthetic complete media [3, 15] and found a correlation between the magnitude of the fitness defect and the number of genetic interactions for each single mutant (genetic interaction degree). The prevailing explanation for this correlation is that genes that display a fitness defect represent the subset that are playing an active role under the condition tested, are additionally not completely buffered by other genes, and/or contribute to a wider variety of cellular processes. We observe a similar correlation between the single mutant fitness defect (as previously measured on synthetic complete media ) and the number of significant condition-specific sensitivities in our study (r = 0.33, P < 5 × 10−100; Pearson). Additionally, there is a partial correlation between the number of genetic interactions a gene has and the number of environments with which it interacts, even after controlling for single mutant fitness defect (r = 0.18, P < 5 × 10−31; Pearson). This echoes a previously observed correlation between genetic interaction degree and sensitivities in more complex chemical environments (r = 0.4, P < 10−5) [6, 15]. These results confirm that our study is uncovering more effects for genes known to be pleiotropic or central under a variety of environmental backgrounds . These findings also suggest that hubs are conserved across different network types, with many of the same genes conferring robustness to genetic, chemical, and environmental perturbations.
Mutant sensitivity profiles are predictive of gene function
Previous genetic interaction studies have shown that high profile similarity for mutant sensitivity across varied environmental conditions or diverse genetic backgrounds (for example, genetic interaction profiles) is highly predictive of similar gene function [5, 7, 15]. We applied an analogous logic to our data to see if similar environmental sensitivity profiles would also be predictive of similar function. Using co-annotation to an informative set of Gene Ontology (GO) terms [16, 17] as our standard for functional similarity, we ranked all pairs of genes by their profile similarity (Pearson) and evaluated these rankings with respect to known functional relationships. We measured a precision of approximately 35% at a recall of 1,000 gene pairs (2-fold over a random baseline of 17%; Figure 3b). Additionally, when we restrict our predictions to those genes with a known involvement in metabolism (663) we see a much higher precision (precision ~ 65% at recall = 100), though a similar performance over the increased background rate (1.7-fold over 38%; see Materials and methods). The higher performance for metabolism-related predictions is likely due to the direct relevance of the environmental conditions chosen to the study of basic metabolism. Thus, we have demonstrated an ability to predict general gene function using the guilt-by-association principle, and the diverse environments chosen for this assay are well-suited to reveal sensitivities in the metabolic network of this newly created prototrophic collection.
Metabolic network models show modest ability to predict experimental data
Prediction of condition-specific slow growth proved consistently above random expectation (Figure 4), though values of precision are much lower than those previously reported in predicting qualitative essentiality (>90% ). One key difference between our study and Snitkin et al.  (as with Dudley et al. in the section on galactose sensitivity above) is the latter's inclusion of antimycin A in the media, which inhibits energy production from respiration, whereas our strains could naturally respire and ferment. Our results show an advantage for the more recent Yeast5 model over the iMM904 model, as well as a slight advantage for standard FBA over MoMA. The Yeast5 model was able to perform above random expectation in 14 out of 20 conditions with a mean precision of 25% and a mean recall of 18% (Figure 4; Additional file 6). Recall scores for MoMA were generally higher than for FBA owing to a much smaller fraction of the predictions equal to the mode, though this was generally associated with a loss of precision. Galactose conditions appear to be well captured by the two models, and consistently perform above random. By contrast, all three conditions for which no model-method achieved significance involved glucose (glucose:allantoin, glucose:glutamine, glucose:urea). Thus, while the overall performance demonstrates an above-random ability of these models to predict quantitative and condition specific perturbation effects, their modest precision and recall scores (<50%) suggest substantial room for improvement.
An examination of false positives (predicted sensitive by the model but not observed in the data) and false negatives (observed sensitive, not predicted) showed some functional coherency. Specifically, Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment of false positives in many conditions revealed connections to central carbon metabolism (for example, the tricarboxylic acid cycle), and half of the conditions showed enrichment for the KEGG sulfur metabolism pathway in the model for false positives (Additional file 6). This suggests potential pathways that may need attention for the development of improved models.
We also attempted to leverage existing metabolic models to demonstrate the widespread metabolic consequences of these common auxotrophies. To accomplish this, we ran the models again using prototrophic and auxotrophic versions of the network on glucose:ammonium and characterized each metabolite as either: i) produced in the auxotroph and the prototroph; ii) produced in the prototroph only; or iii) included in the model but not produced in an optimal solution (see Materials and methods). The simulations show that a significant proportion of producible metabolites (18% in iMM904 and 7% in Yeast5; Figure 4c) are unavailable in the auxotrophic network. This means that consequences of using auxotrophic strains, even under supplementation for their specific deficiencies, may have a broader impact than expected. It is our hope that the collection and accompanying growth data presented here will prove invaluable to the metabolic modeling community as it continues to refine the structure of its models as well as their underlying biological assumptions.
Broad environmental surveys address incomplete gene annotations
A primary motivation for measuring fitness across diverse environments is the discovery of novel phenotypes for mutants that have near wild-type fitness under previously tested conditions. The existence of such mutants in a eukaryotic genome with approximately 6,000 genes is driven by two main factors. The first is genetic redundancy, whereby genes are performing vital functions within the cell, but their importance is not captured by single mutant phenotypes because other genes are present that buffer the loss of function. This occurs at both the level of individual genes buffering one another (for example, duplicate genes [23, 24]) and the level of larger network structures (for example, parallel pathways). These buffered functions are rapidly being mapped by genetic interaction studies that delete multiple genes simultaneously [2, 4, 11, 15, 25, 26]. The remaining contributing factor is environmental robustness, whereby a gene presumably has an important function under some evolutionarily relevant circumstance that is not reflected in a laboratory environment (for example, nutrients/media, temperature, stress). Thus, an important motivation for complete pairwise coverage of basic metabolic conditions is the detection of novel fitness defects for genes that become necessary only as the condition space is more broadly surveyed. Interestingly, of the 729 remaining uncharacterized mutants in the auxotrophic collection for which we have single mutant fitness measurements in synthetic complete media, a significant fraction of them (609) have a fitness greater than 99% of wild-type (hypergeometric P < 7 × 10−66) . Despite the ever-increasing availability of high-throughput genomic data for these genes, the task of eliminating this set has seen only marginal success since 2007 . It is possible that these genes (many of which only have orthologs in other yeasts) may be responsible for functions needed in the native environment of yeast but unnecessary under standard laboratory conditions. Still others may be required in the lab, but only after varying the nutrient conditions. The focus of recent chemical genomics work on subjecting yeast to an extremely broad range of chemical environments is helping to address these genes [5, 6], but auxotrophy in the deletion collection had precluded measurements of growth on simple but directly relevant metabolic conditions. Here we address the potential impact of these data on both uncharacterized genes and genes of little phenotypic consequence in standard conditions.
Novel effects for genes with high fitness in standard conditions
Novel phenotypes for uncharacterized ORFs
Approximately 13% of the S. cerevisiae deletion collection is composed of uncharacterized ORFs , 692 of which are included in this study. Nearly 25% of these uncharacterized genes show a significant effect in two or more non-glucose:ammonia conditions (172/692; Figure 5a) compared to the 4% expected given our FDR.
One such example with a very specific nitrogen sensitivity signature is FMP32. The fmp32Δ0 strain displays dramatically decreased fitness under arginine and proline conditions. While the protein product of FMP32 has been detected in highly purified mitochondria , the gene is otherwise uncharacterized. The fmp32Δ0 strain was included in our liquid confirmation assay and these sensitivities were confirmed in this independent, small-scale assay (Figure 5b). This highly specific signature appears to be completely unique to the fmp32Δ0 strain, as no other mutant in the collection shows a similar sensitivity profile.
The genes with the highest profile similarity to FMP32 are PUT1, PUT3, and RRF1, which have been previously implicated in proline utilization (PUT1, PUT3) and mitochondrial ribosome recycling/mitochondrial protein synthesis during respiration (RRF1[32, 33]). PUT3 induces PUT1 transcription when proline is present as the best available nitrogen source and the latter (along with PUT2) is responsible for the conversion of proline into glutamate for further use as a nitrogen source. Our analysis suggests that FMP32 is similarly involved in the respiratory response under proline, though the reason for its additional sensitivity under arginine remains unclear. These examples show the utility of interactions between genes and simple environments in uncovering the function of both individual uncharacterized genes and genes without a previously observed fitness defect in more complete media.
Clustering of metabolic conditions reveals carbon source as primary factor driving mutant profiles
Matrix factorization distinguishes carbon from nitrogen effects
Further examination of gene and environmental profiles after clustering revealed cases where a gene (for example, FMP32) exhibited an effect in multiple instances of a particular nitrogen source (for example, proline or arginine), but without a specific pattern with regard to carbon source (or vice versa). This is expected behavior for genes required for the utilization of a particular carbon/nitrogen source regardless of the context. In order to more formally extract a list of sensitivities for each source of carbon or nitrogen regardless of its partner, we employed a method known as non-negative matrix factorization (NMF) [34, 35] to decompose our experimental data into a collection of characteristic source signatures. When a matrix of these source signatures is multiplied by a matrix describing the source composition in each of our conditions, the result should approximate our experimental observations. NMF allows us to run this multiplication in reverse and fit the source signatures as an unknown factor. These source signatures are available in Additional file 7, and several of them demonstrate enrichment for related GO terms and KEGG pathways.
One example of a decomposed signature involves genes that are sensitive when glutamate is chosen as a nitrogen source. These genes are enriched for annotations relating to endocytosis, endosome and vacuole related transport, and retrograde transport (Additional file 7). Extracellular glutamate decreases cellular amino acid permease activity by redirecting intracellular trafficking of the permease Gap1 from the plasma membrane to the vacuolar membrane . Many of the mutations in our glutamate signature increase Gap1 activity by misdirecting the protein to the plasma membrane . Although GAP1 is transcribed at equal levels in cells grown on urea and glutamate, permease activity in urea grown cells is 100 times higher than glutamate-grown cells . Inappropriate Gap1 activity is toxic in the context of high concentrations of single amino acids , and we speculate that the inappropriate trafficking in these mutants causes high levels of permease activity that inhibit cell growth.
Many mutants (92) appear in both the galactose and ribose signatures, and overlapping GO enrichments in these conditions reveal many of these genes to have known involvement in various aspects of respiration. For example, enrichment for GO terms relating to mitochondrial organization and translation, as well as 'aerobic respiration', appear highly significant in both of these signatures (Additional file 7). Exceptions include GAL pathway mutants that fall uniquely into the galactose carbon signature ('galactose metabolic process' P < 1.3 × 10−4) and genes involved in acetyl-CoA biosynthesis that appear to be specifically sensitive under ribose (P < 1.4 × 10−6). As more complex environments are mapped, multivariate statistical techniques will become increasingly important in determining which environmental constituents are actually relevant to which experimental observations, and care should be taken when designing experiments to ensure their successful application (for example, complete combinatorial coverage of relevant environmental factors).
Environmental and genetic perturbations can provoke similar cellular states
The idea of comparing environmental and genetic perturbations can be generalized to other genome-wide perturbation data as well. For example, we observe significant correlations between our glutamate signature and a rapamycin sensitivity profile as measured by two different chemical genomic screens (Hillenmyer et al. P < 10−18; Parsons et al. P < 10−9). The enrichment for transport-related terms observed in the glutamate signature (above), and its similarity to a rapamycin profile make sense given that rapamycin redirects trafficking of Gap1 from the plasma membrane to the vacuole . Thus, the same set of mutations in vesicle trafficking that lead to inappropriate expression of Gap1 permease activity in cells grown on glutamate also cause inappropriate permease activity following rapamycin treatment.
The creation of the original yeast deletion collection has had a profound impact on the way in which reverse genetic experiments are performed. Yet despite a staggering number of successful studies, the inherent auxotrophies create a major blind-spot in a fundamental area of cellular function, and previous reviews of the topic have called for the creation and use of standardized prototrophic strains for metabolic experiments . Recently, Mülleder and colleagues  have addressed the deletion collection auxotrophies by introducing a plasmid containing sequences for HIS3, URA3, LEU2, and MET15. The resource used in this study differs in that URA3, LEU2, and MET15 are in their native genomic locations, with the exception of HIS3, which is provided by Schizosaccharomyces pombe HIS5 under the SGA reporter . Without the necessity for plasmid selection, or possible effects on gene expression due to non-chromosomal location, we anticipate that our deletion collection will see frequent use by experimentalists.
The use of a genome-wide prototrophic strain collection enables truly informative sensitivity screening in metabolically controlled conditions. This represents a first step in probing how nutrients in the environment jointly affect cellular response with or without additional genetic perturbation. This study demonstrates that much work is yet to be done to understand growth in even simple environments. A solid grasp of the surprisingly complex responses to simple environments will add much needed context to studies done in more complex environments.
This study has demonstrated the potential of this collection, when screened against simple environments, to uncover phenotypes for hundreds of mutants that are phenotypically normal in standard lab conditions. We believe that the stock of simple experiments that might reveal a phenotype for these mutants has not yet been exhausted and expect that this whole-genome prototrophic collection will be an invaluable resource to the community. The rising number of metabolomics studies, fueled in part by the increasing accuracy of experimental mass-spectrometry, as well as the growing interest in metabolism as central to many common ailments in humans, make it more important than ever to properly design metabolically relevant experiments in the model eukaryote S. cerevisiae. Central to that goal is a version of the deletion collection that is unhindered by historical auxotrophic requirements.
For example, while central metabolism is unrivaled among cellular processes with respect to our ability to make in silico predictions from constraint-based metabolic models, it is far from a fully understood system. Our results show a generally weak ability to predict condition-specific sensitivities, though performance is clearly above a random baseline. The prediction of condition-specific sensitivities is admittedly more difficult than the prediction of sensitivities in general, but it was our estimation that FBA and MoMA would be well suited to approximate our observations given our simple experimental setup. Their only moderate success in doing so demonstrates the current limitations of constraint-based modeling and the difficulty of relating models built from biomass predictions to quantitative growth rate data. There might be several possible reasons for the discrepancy between in silico and in vivo results. First, the success of predicting growth defects hinges on the proper formulation of biomass composition. While a single biomass composition is used for all our simulations, it likely changes across environmental conditions. Future studies could address this issue by measuring the composition of yeast cells under different nutrient settings. A second limitation of purely flux-based models is their inability to make predictions about components that have an indirect effect on metabolism. Consider, for example, the enrichment for transport-related genes whose deletion confers glutamate-specific sensitivities. Their putative role in nutrient sensing and signaling reflects the fact that, despite its constrained nature, the metabolic network operates as part of a much larger and more dynamic network. More generally, the basic constraint-based modeling approaches ignore regulatory mechanisms. Several attempts have been made to bridge this gap and they rely either on 'omic' data to constrain the activity of specific reactions [42–44] or on integrating a mathematical representation of gene regulation with the metabolic model [45–47]. We feel that the availability of this whole-genome collection and accompanying growth data well suited to studies of metabolism will help the community to develop and test novel models and methods to better capture the operation of the greater cellular network.
Central to the understanding of the network as a whole is the idea that a whole-genome screen reveals indirect as well as direct consequences of the perturbation tested. Positive gene-environment interactions under ribose conditions may well illustrate this point. The median z-score for the 166 genes annotated to 'chromosome segregation' in GO is negative for all seven galactose conditions, yet positive for all seven ribose conditions (binomial sign-test P < 6.2 × 10−5). We believe this shift may be explained by fundamental cellular rate limitations. Failure to segregate chromosomes in the midst of even moderate growth (for example, galactose) can have very severe consequences, ultimately limiting growth rate, whereas comparatively slow growth (for example, ribose) affords additional time for slowly segregating mutants to complete segregation. These mutants grow faster than we expect despite no apparent link between carbon metabolism and chromosome segregation. Thus, growth rates under one condition disclose information about the interplay between a wide variety of cellular subsystems, giving us a readout of the internal cellular state. Similarly, a mutant profile across many environments gives us information about how essential that gene may be in any of those various cellular states, in addition to elucidating any direct role that gene may have in direct utilization of the provided nutrients. Analysis of our growth data recapitulated the role of vesicle trafficking in the regulation of the amino acid permease Gap1, relating growth on glutamate to the drug rapamycin. This broader view of whole-genome screen information then allows for integration of profiles across different perturbation types (chemical, genetic, environmental), and should ultimately aid us in applying knowledge gained in one arena to observations made in another.
Materials and methods
Construction of a prototrophic deletion collection
As recently described , the strains in the standard MAT a deletion collection (MAT a yfgΔ0::KanMX his3Δ1 leu2Δ0 met15Δ0 ura3Δ0)  were mated to a MATα can1Δ::STE2pr-SpHIS5 his3Δ1 lyp1Δ0 strain, creating diploids (selection on minimal media + his + G418). These were sporulated and successive pinnings on selective media were used to select prototrophic MAT a strains carrying each deletion allele. These prototrophic strains were organized into an array of 16 plates including one entire plate of the wild-type strain (hoΔ::KanMX), with additional wild-type replicates in each row and column of every plate (701 in all). The entire prototrophic collection is available upon request, as is the individual SGA-ready prototroph strain for crossing into other collections.
Minimal growth media were prepared using yeast nitrogen base (BD Difco, Sparks, Maryland,) with the specified carbon and nitrogen sources. Carbon sources included glucose, galactose, ribose, and glycerol. Nitrogen sources included ammonium, allantoin, arginine, glutamate, glutamine, proline, and urea. Carbon sources were provided at a concentration of 2%; nitrogen sources were 3.8 mM with respect to nitrogen.
Calculation of growth rate
Sixteen 16 × 24 well plates were grown in 28 chemical conditions for 24 to 48 hours. Plates were scanned on a flatbed transparency scanner at 0, 5, 10, 24 and, in the case of glycerol, 48 hours. Each condition is composed of one carbon source and one nitrogen source. In total, 4,772 mutants were grown, and colony areas were extracted from tiff images by CellProfiler  and precise time points were taken from EXIF data in the digital images. These values were used to compute an estimate of the growth rate of each colony equal to the slope of the least-squares linear fit of area (pixels) to time (seconds). Colonies with insufficient data were given a growth rate of NaN, colonies with a negative calculated growth rate were defined to have a growth rate of 0.
Definition and construction of a reference condition
The result of this approach is quite robust to the choice of PlateA, and so we used whichever replicate had the fewest number of missing values and would therefore provide the most complete LOWESS fit. After normalizing five replicates to the sixth, all six were averaged together to create one reference plate, and this procedure is repeated 16 times to create a glucose:ammonium reference for each array plate.
Normalization of experimental rates against reference
Recovery of missing data
In certain cases, a growth rate of NaN was assigned to a colony due to insufficient data being collected by CellProfiler. In an effort to recover any good data, these cases were visually inspected by five researchers operating independently and a vote was taken to determine whether to leave it as missing data (NaN) or assign it a growth rate of 0, indicating that the colony appeared to be correctly plated but non-viable. In total, 1,362 of 2,601 colonies were recovered this way.
Transformation from normalized rates to z-scores
For each array plate, at each position, a strain-wise standard deviation is calculated across the residuals of the six glucose:ammonium (GA) replicates.
These z-scores are an expression of the difference in magnitude and direction between the growth observed at each position of a plate under a given condition from the same position (and hence deletion) under the reference GA model.
Spatial smoothing procedure
The plate level spatial smoothing filter is similar to that found in . First, temporarily replace any extreme values (top and bottom 5%) along with NaNs with the plate mean. Second, replace previous NaN positions with values from a two-dimensional symmetric gaussian filter. Third, compute and subtract the residual between the two-dimensional smoothed plate and its mean.
Choosing effect thresholds
Each condition had 701 wild-type replicates. The mean and standard deviation of the set of wild-type z-scores were used to define a normal distribution against which P-values for the experimental z-scores could be calculated. This information allowed the use of Benjamini-Hochberg procedure to establish condition-specific effect thresholds as a function of a desired FDR (Additional file 4).
Liquid growth confirmation assay
The growth rate of 40 mutants in a liquid growth assay was measured across 20 of the experimental conditions excluding ribose:arginine and all glycerol pairings. Liquid culture assays were not performed for the ribose:arginine conditions because the combination of these carbon and nitrogen sources did not allow arginine to maintain adequate solubility over the duration of the experiment. The precipitation of arginine prevented accurate optical density readings from being obtained and thus these data were excluded from our subsequent analyses. Six replicate wells contained the wild-type strain and each mutant strain was represented twice. Cells were pre-grown on glucose:ammonia medium and diluted at a low density into the growth medium of interest. Growth rates were determined as the maximum optical density (saturation) divided by the time to saturation. A simple model was favored in order to robustly accommodate drastic differences in curve characteristics between fast growth and slow growth conditions (for example, galactose versus ribose).
We adjusted the liquid growth scores by dividing the mean of mutant growth slopes by the mean of wild-type growth slopes in the relevant condition. We further normalized these scores by dividing them by the corresponding adjusted mutant score in glucose:ammonium so they would reflect condition-specific effects, similar to our modified z-score derived from the agar experiment.
Gene Ontology and KEGG enrichments and co-annotation standards
Genome-scale metabolic modeling (FBA and MoMA)
Two S. cerevisiae metabolic models were used for mutant biomass prediction. The Yeast Consensus Reconstruction version 5.35 (Yeast5)  and iMM904 . Yeast5 consisted of 898 ORFs, 2,031 reactions and 1,594 metabolites and the iMM904 model contained 901 ORFs, 1,597 reactions and 1,234 metabolites. Default biomass descriptions were used for both models.
Wild-type biomass production flux for each condition was obtained using FBA  in MATLAB with the COBRA Toolbox , which assumes optimal biomass production (that is, maximum biomass yield). Mutant biomass flux was predicted using both FBA  in MATLAB with the COBRA Toolbox  and MoMA  in MATLAB with the ILOG CPLEX optimization suite. MoMA was formulated as a quadratic programming problem, whereby mutant fluxes were selected that minimized the Euclidean distance from an optimal wild-type flux distribution. The yeast wild-type flux distribution was calculated as a network flux solution producing maximal biomass flux, determined by FBA, with minimal total fluxes .
Prediction of positive z-scores was also carried out, though performance was generally below random expectation (Additional file 6). This is likely due to the fact that many positive z-scores corresponded to raw growth rates for mutants that were faster than wild-type under the same condition, a consequence that FBA- and MoMA-based methods would find difficult or impossible to predict.
To calculate the effect of gene deletions on the metabolic network (Figure 4c), sets of producible metabolites were calculated for the complete model, and for a mutant with all four auxotrophic marker genes deleted. Producible metabolites were calculated for both iMM904 and Yeast5 models in the glucose:ammonium media condition by adding a special exchange reaction for each metabolite and iteratively optimizing flux exported through that reaction. If the export flux for a given metabolite exceeded 0.001 (with an upper and lower bound on internal reactions set to ±1,000), it was classified as 'producible.' A non-zero threshold is required to limit false positives as a result of numerical errors. The threshold was determined to be robust by scaling the upper and lower bounds, as well as the threshold by a large constant and counting the number of producible metabolites. Obtaining consistent results in these experiments led us to conclude that numerical errors are an order of magnitude smaller than contributions from stoichiometry.
Source signature decomposition via modified non-negative matrix factorization
To ensure linear independence among the columns of the Coefficient matrix, we removed all but one glucose:ammonium (glucose:ammonium01) column, removing the same columns in the Data matrix. Traditional NMF would use a multiplicative update algorithm applied to both the Signature and the Coefficient matrix to find the best fit to the data; however, we chose to fix the Coefficient matrix at the initial defined values (0 or 1). This gives each Signature column equal weight and prevents over-fitting caused by the sparsity of the Data matrix and the dramatically different number of non-zero elements from one column to the next. The multiplicative update was applied for 20 iterations, though in practice the results converged in fewer than 10, and repeated trials from different random initializations of the Signature matrix showed the results to be quite stable. Genes were added to the signature list in Additional file 7 if their value exceeded 0.4.
Comparison to SGA data
For the comparison to auxotrophic SGA data represented in Figure 7, the SGA data were taken from . The SGA data and the z-score data (Additional file 3) were independently normalized so that row and column vectors had a euclidean length approximately equal to 1, and missing values were set to 0. Inner product was then used to measure the similarity between SGA 'queries' and environmental profiles. The top 10% of queries in each condition were checked for enrichment for GO terms and KEGG pathway annotations, and the resulting P-values were Bonferroni corrected to account for the number of terms/pathways tested against.
Comparison with previous whole-genome screens on galactose
Figure 3a uses a publicly available image for the basis of the Venn diagram. This image is used with permission under the terms of the Creative Commons Attribution-Share Alike 3.0 Unported license .
flux balance analysis
false discovery rate
Kyoto Encyclopedia of Genes and Genomes
minimization of metabolic adjustment
non-negative matrix factorization
open reading frame
synthetic genetic array.
BJV and CLM are partially supported by grants from the National Science Foundation (DBI 0953881) and the National Institutes of Health (R01HG005084). BJV was also partially supported by the University of Minnesota Doctoral Dissertation Fellowship. DCH is supported by the National Institutes of Health (R01 GM101091-01) and the National Science Foundation (11122240). BP was supported by the Wellcome Trust and the ‘Lendület Program’ of the Hungarian Academy of Sciences. BS was supported by the European Union and the State of Hungary, co-financed by the European Social Fund in the framework of TÁMOP 4.2.4. A/2-11-1-2012-0001 'National Excellence Program'. CN acknowledges funding from the CCSRI, grant number 20830. AAC is the Canada Research Chair in Metabolomics for Enzyme Discovery and is supported by the Ontario Early Researcher Award, by the Canadian Institutes for Health Research, and by the Natural Sciences and Engineering Research Council of Canada, and by the Canadian Foundation for Innovation and the Ontario Leader's Opportunity Fund. DCH, CLM, OGT, and AAC were supported by the National Institute of General Medical Sciences (NIGMS) Center of Excellence P50 GM071508. Finally, we thank David Botstein for advice and insight throughout the project.
- Giaever G, Chu AM, Ni L, Connelly C, Riles L, Véronneau S, Dow S, Lucau-Danila A, Anderson K, André B, Arkin AP, Astromoff A, El-Bakkoury M, Bangham R, Benito R, Brachat S, Campanaro S, Curtiss M, Davis K, Deutschbauer A, Entian K-D, Flaherty P, Foury F, Garfinkel DJ, Gerstein M, Gotte D, Güldener U, Hegemann JH, Hempel S, Herman Z, et al: Functional profiling of the Saccharomyces cerevisiae genome. Nature. 2002, 418: 387-391. 10.1038/nature00935.View ArticlePubMedGoogle Scholar
- Tong AHY, Lesage G, Bader GD, Ding H, Xu H, Xin X, Young J, Berriz GF, Brost RL, Chang M, Chen Y, Cheng X, Chua G, Friesen H, Goldberg DS, Haynes J, Humphries C, He G, Hussein S, Ke L, Krogan N, Li Z, Levinson JN, Lu H, Ménard P, Munyana C, Parsons AB, Ryan O, Tonikian R, Roberts T, et al: Global mapping of the yeast genetic interaction network. Science. 2004, 303: 808-813. 10.1126/science.1091317.View ArticlePubMedGoogle Scholar
- Baryshnikova A, Costanzo M, Kim Y, Ding H, Koh J, Toufighi K, Youn J-Y, Ou J, San Luis B-J, Bandyopadhyay S, Hibbs M, Hess D, Gingras A-C, Bader GD, Troyanskaya OG, Brown GW, Andrews B, Boone C, Myers CL: Quantitative analysis of fitness and genetic interactions in yeast on a genome scale. Nat Methods. 2010, 7: 1017-1024. 10.1038/nmeth.1534.View ArticlePubMedPubMed CentralGoogle Scholar
- Bandyopadhyay S, Mehta M, Kuo D, Sung M-K, Chuang R, Jaehnig EJ, Bodenmiller B, Licon K, Copeland W, Shales M, Fiedler D, Dutkowski J, Guénolé A, van Attikum H, Shokat KM, Kolodner RD, Huh W-K, Aebersold R, Keogh M-C, Krogan NJ, Ideker T: Rewiring of genetic networks in response to DNA damage. Science. 2010, 330: 1385-1389. 10.1126/science.1195618.View ArticlePubMedPubMed CentralGoogle Scholar
- Parsons AB, Brost RL, Ding H, Li Z, Zhang C, Sheikh B, Brown GW, Kane PM, Hughes TR, Boone C: Integration of chemical-genetic and genetic interaction data links bioactive compounds to cellular target pathways. Nat Biotechnol. 2004, 22: 62-69. 10.1038/nbt919.View ArticlePubMedGoogle Scholar
- Hillenmeyer ME, Fung E, Wildenhain J, Pierce SE, Hoon S, Lee W, Proctor M, St Onge RP, Tyers M, Koller D, Altman RB, Davis RW, Nislow C, Giaever G: The chemical genomic portrait of yeast: uncovering a phenotype for all genes. Science. 2008, 320: 362-365. 10.1126/science.1150021.View ArticlePubMedPubMed CentralGoogle Scholar
- Dudley AM, Janse DM, Tanay A, Shamir R, Church GM: A global view of pleiotropy and phenotypically derived gene function in yeast. Mol Syst Biol. 2005, 1: 2005.0001-View ArticlePubMedPubMed CentralGoogle Scholar
- Price ND, Reed JL, Palsson BØ: Genome-scale models of microbial cells: evaluating the consequences of constraints. Nat Rev Microbiol. 2004, 2: 886-897. 10.1038/nrmicro1023.View ArticlePubMedGoogle Scholar
- Pronk JT: Auxotrophic yeast strains in fundamental and applied research. Appl Environ Microbiol. 2002, 68: 2095-2100. 10.1128/AEM.68.5.2095-2100.2002.View ArticlePubMedPubMed CentralGoogle Scholar
- Brem RB, Yvert G, Clinton R, Kruglyak L: Genetic dissection of transcriptional regulation in budding yeast. Science. 2002, 296: 752-755. 10.1126/science.1069516.View ArticlePubMedGoogle Scholar
- Tong AHY, Evangelista M, Parsons AB, Xu H, Bader GD, Pagé N, Robinson M, Raghibizadeh S, Hogue CW, Bussey H, Andrews B, Tyers M, Boone C: Systematic genetic analysis with ordered arrays of yeast deletion mutants. Science. 2001, 294: 2364-2368. 10.1126/science.1065810.View ArticlePubMedGoogle Scholar
- Tong AHY, Boone C: Synthetic genetic array analysis in Saccharomyces cerevisiae. Methods Mol Biol. 2006, 313: 171-192.PubMedGoogle Scholar
- Kuepfer L, Sauer U, Blank LM: Metabolic functions of duplicate genes in Saccharomyces cerevisiae. Genome Res. 2005, 15: 1421-1430. 10.1101/gr.3992505.View ArticlePubMedPubMed CentralGoogle Scholar
- Barford JP, Hall RJ: An examination of the crabtree effect in Saccharomyces cerevisiae: the role of respiratory adaptation. J Gen Microbiol. 1979, 114: 267-275. 10.1099/00221287-114-2-267.View ArticleGoogle Scholar
- Costanzo M, Baryshnikova A, Bellay J, Kim Y, Spear ED, Sevier CS, Ding H, Koh JLY, Toufighi K, Mostafavi S, Prinz J, St Onge RP, VanderSluis B, Makhnevych T, Vizeacoumar FJ, Alizadeh S, Bahr S, Brost RL, Chen Y, Cokol M, Deshpande R, Li Z, Lin Z-Y, Liang W, Marback M, Paw J, San Luis B-J, Shuteriqi E, Tong AHY, van Dyk N, et al: The genetic landscape of a cell. Science. 2010, 327: 425-431. 10.1126/science.1180823.View ArticlePubMedGoogle Scholar
- Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT, Harris MA, Hill DP, Issel-Tarver L, Kasarskis A, Lewis S, Matese JC, Richardson JE, Ringwald M, Rubin GM, Sherlock G: Gene ontology: tool for the unification of biology, The Gene Ontology Consortium. Nat Genet. 2000, 25: 25-29. 10.1038/75556.View ArticlePubMedPubMed CentralGoogle Scholar
- Myers CL, Barrett DR, Hibbs MA, Huttenhower C, Troyanskaya OG: Finding function: evaluation methods for functional genomic data. BMC Genomics. 2006, 7: 187-10.1186/1471-2164-7-187.View ArticlePubMedPubMed CentralGoogle Scholar
- Snitkin ES, Dudley AM, Janse DM, Wong K, Church GM, Segrè D: Model-driven analysis of experimentally determined growth phenotypes for 465 yeast gene deletion mutants under 16 different conditions. Genome Biol. 2008, 9: R140-10.1186/gb-2008-9-9-r140.View ArticlePubMedPubMed CentralGoogle Scholar
- Heavner BD, Smallbone K, Barker B, Mendes P, Walker LP: Yeast 5 - an expanded reconstruction of the Saccharomyces cerevisiae metabolic network. BMC Syst Biol. 2012, 6: 55-10.1186/1752-0509-6-55.View ArticlePubMedPubMed CentralGoogle Scholar
- Zomorrodi AAR, Maranas CCD: Improving the iMM904 S. cerevisiae metabolic model using essentiality and synthetic lethality data. BMC Syst Biol. 2010, 4: 178-10.1186/1752-0509-4-178.View ArticlePubMedPubMed CentralGoogle Scholar
- Orth JD, Thiele I, Palsson BØ: What is flux balance analysis?. Nat Biotechnol. 2010, 28: 245-248. 10.1038/nbt.1614.View ArticlePubMedPubMed CentralGoogle Scholar
- Segrè D, Vitkup D, Church GM: Analysis of optimality in natural and perturbed metabolic networks. Proc Natl Acad Sci U S A. 2002, 99: 15112-15117. 10.1073/pnas.232349399.View ArticlePubMedPubMed CentralGoogle Scholar
- VanderSluis B, Bellay J, Musso G, Costanzo M, Papp B, Vizeacoumar FJ, Baryshnikova A, Andrews B, Boone C, Myers CL: Genetic interactions reveal the evolutionary trajectories of duplicate genes. Mol Syst Biol. 2010, 6: 429-View ArticlePubMedPubMed CentralGoogle Scholar
- Musso G, Costanzo M, Huangfu M, Smith AM, Paw J, San Luis B-J, Boone C, Giaever G, Nislow C, Emili A, Zhang Z: The extensive and condition-dependent nature of epistasis among whole-genome duplicates in yeast. Genome Res. 2008, 18: 1092-1099. 10.1101/gr.076174.108.View ArticlePubMedPubMed CentralGoogle Scholar
- Ryan CJ, Roguev A, Patrick K, Xu J, Jahari H, Tong Z, Beltrao P, Shales M, Qu H, Collins SR, Kliegman JI, Jiang L, Kuo D, Tosti E, Kim H-S, Edelmann W, Keogh M-C, Greene D, Tang C, Cunningham P, Shokat KM, Cagney G, Svensson JP, Guthrie C, Espenshade PJ, Ideker T, Krogan NJ: Hierarchical modularity and the evolution of genetic interactomes across species. Mol Cell. 2012, 46: 691-704. 10.1016/j.molcel.2012.05.028.View ArticlePubMedPubMed CentralGoogle Scholar
- Schuldiner M, Collins SR, Thompson NJ, Denic V, Bhamidipati A, Punna T, Ihmels J, Andrews B, Boone C, Greenblatt JF, Weissman JS, Krogan NJ: Exploration of the function and organization of the yeast early secretory pathway through an epistatic miniarray profile. Cell. 2005, 123: 507-519. 10.1016/j.cell.2005.08.031.View ArticlePubMedGoogle Scholar
- Cherry JM, Hong EL, Amundsen C, Balakrishnan R, Binkley G, Chan ET, Christie KR, Costanzo MC, Dwight SS, Engel SR, Fisk DG, Hirschman JE, Hitz BC, Karra K, Krieger CJ, Miyasato SR, Nash RS, Park J, Skrzypek MS, Simison M, Weng S, Wong ED: Saccharomyces Genome Database: the genomics resource of budding yeast. Nucleic Acids Res. 2012, 40: D700-D705. 10.1093/nar/gkr1029.View ArticlePubMedGoogle Scholar
- Peña-Castillo L, Hughes TR: Why are there still over 1000 uncharacterized yeast genes?. Genetics. 2007, 176: 7-14. 10.1534/genetics.107.074468.View ArticlePubMedPubMed CentralGoogle Scholar
- Landry CR, Oh J, Hartl DL, Cavalieri D: Genome-wide scan reveals that genetic variation for transcriptional plasticity in yeast is biased towards multi-copy and dispensable genes. Gene. 2006, 366: 343-351. 10.1016/j.gene.2005.10.042.View ArticlePubMedGoogle Scholar
- Reinders J, Zahedi RP, Pfanner N, Meisinger C, Sickmann A: Toward the complete yeast mitochondrial proteome: multidimensional separation techniques for mitochondrial proteomics. J Proteome Res. 2006, 5: 1543-1554. 10.1021/pr050477f.View ArticlePubMedGoogle Scholar
- Huang HL, Brandriss MC: The regulator of the yeast proline utilization pathway is differentially phosphorylated in response to the quality of the nitrogen source. Mol Cell Biol. 2000, 20: 892-899. 10.1128/MCB.20.3.892-899.2000.View ArticlePubMedPubMed CentralGoogle Scholar
- Teyssier E, Hirokawa G, Tretiakova A, Jameson B, Kaji A, Kaji H: Temperature-sensitive mutation in yeast mitochondrial ribosome recycling factor (RRF). Nucleic Acids Res. 2003, 31: 4218-4226. 10.1093/nar/gkg449.View ArticlePubMedPubMed CentralGoogle Scholar
- Kanai T, Takeshita S, Atomi H, Umemura K, Ueda M, Tanaka A: A regulatory factor, Fil1p, involved in derepression of the isocitrate lyase gene in Saccharomyces cerevisiae–a possible mitochondrial protein necessary for protein synthesis in mitochondria. Eur J Biochem. 1998, 256: 212-220. 10.1046/j.1432-1327.1998.2560212.x.View ArticlePubMedGoogle Scholar
- Lee DD, Seung HS: Learning the parts of objects by non-negative matrix factorization. Nature. 1999, 401: 788-791. 10.1038/44565.View ArticlePubMedGoogle Scholar
- Lee DD, Seung HS: Algorithms for non-negative matrix factorization. Adv Neural Inf Process Syst. 2001, 13: 556-562.Google Scholar
- Roberg KJ: Physiological regulation of membrane protein sorting late in the secretory pathway of Saccharomyces cerevisiae. J Cell Biol. 1997, 137: 1469-1482. 10.1083/jcb.137.7.1469.View ArticlePubMedPubMed CentralGoogle Scholar
- Chen EJ, Kaiser CA: Amino acids regulate the intracellular trafficking of the general amino acid permease of Saccharomyces cerevisiae. Proc Natl Acad Sci U S A. 2002, 99: 14837-14842. 10.1073/pnas.232591899.View ArticlePubMedPubMed CentralGoogle Scholar
- Stanbrough M, Magasanik B: Transcriptional and posttranslational regulation of the general amino acid permease of Saccharomyces cerevisiae. J Bacteriol. 1995, 177: 94-102.PubMedPubMed CentralGoogle Scholar
- Risinger AL, Cain NE, Chen EJ, Kaiser CA: Activity-dependent reversible inactivation of the general amino acid permease. Mol Biol Cell. 2006, 17: 4411-4419. 10.1091/mbc.E06-06-0506.View ArticlePubMedPubMed CentralGoogle Scholar
- Chen EJ, Kaiser CA: LST8 negatively regulates amino acid biosynthesis as a component of the TOR pathway. J Cell Biol. 2003, 161: 333-347. 10.1083/jcb.200210141.View ArticlePubMedPubMed CentralGoogle Scholar
- Mülleder M, Capuano F, Pir P, Christen S, Sauer U, Oliver SG, Ralser M: A prototrophic deletion mutant collection for yeast metabolomics and systems biology. Nat Biotechnol. 2012, 30: 1176-1178. 10.1038/nbt.2442.View ArticlePubMedPubMed CentralGoogle Scholar
- Becker SA, Palsson BO: Context-specific metabolic networks are consistent with experiments. PLoS Comput Biol. 2008, 4: e1000082-10.1371/journal.pcbi.1000082.View ArticlePubMedPubMed CentralGoogle Scholar
- Shlomi T, Cabili MN, Herrgård MJ, Palsson BØ, Ruppin E: Network-based prediction of human tissue-specific metabolism. Nat Biotechnol. 2008, 26: 1003-1010. 10.1038/nbt.1487.View ArticlePubMedGoogle Scholar
- Jensen PA, Papin JA: Functional integration of a metabolic network model and expression data without arbitrary thresholding. Bioinformatics. 2011, 27: 541-547. 10.1093/bioinformatics/btq702.View ArticlePubMedGoogle Scholar
- Covert MW, Schilling CH, Palsson B: Regulation of gene expression in flux balance models of metabolism. J Theor Biol. 2001, 213: 73-88. 10.1006/jtbi.2001.2405.View ArticlePubMedGoogle Scholar
- Covert MW, Xiao N, Chen TJ, Karr JR: Integrating metabolic, transcriptional regulatory and signal transduction models in Escherichia coli. Bioinformatics. 2008, 24: 2044-2050. 10.1093/bioinformatics/btn352.View ArticlePubMedGoogle Scholar
- Chandrasekaran S, Price ND: Probabilistic integrative modeling of genome-scale metabolic and regulatory networks in Escherichia coli and Mycobacterium tuberculosis. Proc Natl Acad Sci U S A. 2010, 107: 17845-17850. 10.1073/pnas.1005139107.View ArticlePubMedPubMed CentralGoogle Scholar
- Gibney PA, Lu C, Caudy AA, Hess DC, Botstein D: Yeast metabolic and signaling genes are required for heat shock survival and have little overlap with the heat-induced genes. Proc Natl Acad Sci U S A. 2013, 110: E4393-E4402. 10.1073/pnas.1318100110.View ArticlePubMedPubMed CentralGoogle Scholar
- Jones TR, Kang IH, Wheeler DB, Lindquist RA, Papallo A, Sabatini DM, Golland P, Carpenter AE: Cell Profiler Analyst: data exploration and analysis software for complex image-based screens. BMC Bioinformatics. 2008, 9: 482-10.1186/1471-2105-9-482.View ArticlePubMedPubMed CentralGoogle Scholar
- Kanehisa M, Goto S, Kawashima S, Okuno Y, Hattori M: The KEGG resource for deciphering the genome. Nucleic Acids Res. 2004, 32: D277-D280. 10.1093/nar/gkh063.View ArticlePubMedPubMed CentralGoogle 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 BØ: Quantitative prediction of cellular metabolism with constraint-based models: the COBRA Toolbox v2.0. Nat Protoc. 2011, 6: 1290-1307. 10.1038/nprot.2011.308.View ArticlePubMedPubMed CentralGoogle Scholar
- Holzhütter H-G: The principle of flux minimization and its application to estimate stationary fluxes in metabolic networks. Eur J Biochem. 2004, 271: 2905-2922. 10.1111/j.1432-1033.2004.04213.x.View ArticlePubMedGoogle Scholar
- Venn’s four ellipse construction. http://en.wikipedia.org/wiki/File:Venn's_four_ellipse_construction.svg,
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 credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.