Plasticity of genetic interactions in metabolic networks of yeast

FBA Reveals a High Frequency of Condition-Dependent Genetic Interactions.

We have extended previous studies (17) by applying FBA to a genome-scale metabolic network model of yeast (S. cerevisiae) to calculate genetic interactions. The previously reconstructed metabolic network (18) consists of 672 genes and 745 unique biochemical reactions and incorporates external nutrients and the corresponding transport processes. The impacts of all possible single- and double-gene deletions were calculated for 53 nutritional environments, including various carbon sources (see SI Materials and Methods). The analysis identified 98 gene pairs that were predicted to be involved in a SL relationship under at least one of the conditions investigated (SI Tables 3 and 4). Only 14.3% of these SL relationships were displayed under all nutrient conditions investigated, and 50% of them are restricted to only one or two nutritional environments (Fig. 3).

The condition-specificity of interactions does not appear to be randomly distributed in the metabolic network. SL interactions between gene pairs annotated to different metabolic subsystems are present in a significantly smaller number of environments than those that are annotated to the same subsystem (Mann–Whitney U test, P < 0.02; because enzymes catalyzing the same reaction, by definition, have the same functional annotation, they were excluded from this analysis). Moreover, more than half (56.3 ± 2.7%) of all genetic interactions remain undetected when only a single environment is investigated. These results not only provide a link between compensation of null mutations and the environment but also suggest that systematic genetic interaction screens (which are generally restricted to a single condition) may miss many of the extant interactions.

Experimental Tests on the Reliability of the Model.

The predictions of the in silico model were tested by in vivo double-gene-deletion experiments (17 cases, Table 1, Materials and Methods) and by extracting published experimental data from the literature (32 cases; see SI Materials and Methods). This procedure enabled us to validate ≈60% of the total number of genetic interactions that we predicted to be present on either minimal or rich media (SI Table 5). Double deletants were constructed by sporulating and dissecting heterozygous diploids from crosses between two single-gene haploid deletants (see SI Materials and Methods). Next, we assessed the viability of double deletants by inspecting growth on plates. In 12 of the 17 cases investigated, we observed a clear synthetic sick or lethal (SSL) phenotype under the predicted growth condition (Table 1). The model also accurately captures changes in the presence of synthetic genetic interactions between media (see below). However, in five cases, the double mutant formed colonies qualitatively indistinguishable from the single-gene deletants.

These apparently false predictions may indicate that the model has only limited resolution. It may be, for example, that FBA accurately predicts the direction, but not the strength, of genetic interaction between genes. To explore whether weak genetic interactions, which are undetectable by a simple plate-growth assay, could be responsible for some of these false predictions, we measured the growth rates of all viable double deletants, and those of the corresponding single deletants, using an established protocol (19) (see Materials and Methods). In two of the five investigated cases, we found evidence for weak (but statistically significant) negative epistasis between the predicted gene pairs (Table 1).

Although only a limited number of interactions were tested experimentally, the results suggest that FBA can reliably detect genetic interactions in the metabolic network of yeast. Future large-scale experimental screens are required to get a precise estimate of the fraction of false-positive and false-negative predictions. As a preliminary to such a larger study, we augmented our experimental results with literature data available on single- and double-deletant strains (see SI Materials and Methods). Overall, we were able to test 49 predicted interactions (SI Table 5) and estimate that ≈49% (24 of 49) of them were correct and that, in 53% of the cases, at least the sign of epistasis was consistent with the predictions. In a similar vein, FBA can identify ≈24% of a curated list (20) of previously described SL interactions between metabolic genes. Both values are at least two orders of magnitude higher than expected by chance (P < 10⁻²⁸⁷, see SI Materials and Methods).

Gene Duplicates Can Explain Many of the False Predictions.

Lack of an observable growth defect in three of the experimentally observed cases could be due to the presence of gene duplicates with redundant functions that are not represented in the current metabolic reconstruction. We investigated this possibility by determining whether one or the other member of the gene pairs investigated had a gene duplicate that might provide compensation for one missing function in the double deletant. One member of the gene pair had a paralog in all three cases. Construction of triple-deletion strains (SI Materials and Methods) revealed strong negative epistasis in all three cases (SI Fig. 5).

For example, CHO2 and CPT1 are erroneously predicted to show a synthetic genetic interaction on rich medium. We hypothesized that this interaction might be masked by EPT1, a duplicate of CPT1. The two encoded proteins show 56% amino acid sequence similarity to each other and have different primary catalytic activities. However, some studies suggest that although Cpt1p accounts for 95% of phosphatidylcholine synthesis in vivo, the remaining 5% is likely to be catalyzed by Ept1p (21). Remarkably, deletion of all three genes simultaneously resulted in a much stronger growth defect than observed for any of the pair-wise deletions (SI Fig. 5).

In addition, our data suggest that interaction between OPI3 and PCT1 is masked by MUQ1, a distant paralog of PCT1 (the products of the two genes share 36% amino acid sequence similarity). The triple opi3Δ/pct1Δ/muq1Δ has a more severe phenotype than either double mutant (SI Fig. 5). Although Pct1p and Muq1p catalyze related reactions, they are generally believed to have different substrate specificities. Further biochemical studies will be needed to confirm whether Muq1p has the catalytic activity necessary to mitigate the effect of the pct1Δ/opi3Δ double deletion.

Detailed investigation of the false predictions can thus be used to generate novel biochemical hypotheses and refine the in silico model. Moreover, these results suggest that even duplicates with low sequence similarities and partly altered functions can compensate null mutations in each other. The importance of appropriate modeling of paralogs/gene duplicates is further underscored by inspection of false-negative predictions (i.e., true pair-wise interactions not predicted by the model). Many of these previously reported interacting gene pairs are predicted to participate in higher-order genetic interactions because of the presence of a gene duplicate with overlapping functions (i.e., an isoenzyme) (see SI Table 6). For example, TDH2 and TDH3 show synthetic lethality in vivo (22); however, our simulations show that tdh2Δ tdh3Δ double mutant is compensated by TDH1, a gene encoding an additional glyceraldehyde-3-phosphate dehydrogenase isoenzyme in the model. Lack of in vivo compensatory capacity of Tdh1p might be explained by its relatively low expression level compared with Tdh2p and Tdh3p (23). Thus, in addition to correctly assigning reactions to paralogous genes, incorporation of regulatory constraints (12) and information on maximum enzyme capacities would also be needed to more accurately model the behavior of isoenzymes. Transcriptional reprogramming upon gene deletion (8) may also have an influence on predicting deletant phenotypes.

Two Explanations for the Condition-Specificity of Genetic Interactions.

We found empirical evidence of condition-specific epistasis for 14 validated SSL gene pairs, of which 11 were correctly predicted (SI Table 7), suggesting that our modeling framework is able to capture variation in the incidence of genetic interactions across a range of environmental conditions. There could be at least two explanations for the condition-dependence of these genetic interactions (see Fig. 1). First, members of the synthetically interacting gene pairs make important individual contributions to growth under different nutritional conditions. Alternatively, the double-deletant strain becomes viable under different conditions. There is experimental evidence for both explanations (Fig. 4, SI Table 7).

CHO2 and PCT1 are genes that encode two enzymes that each catalyze a step in two different pathways responsible for phosphatidylcholine synthesis (the phosphatidylethanolamine methylation pathway and the Kennedy pathway, respectively; see Fig. 4A). In agreement with our first explanation of condition-dependence, we find that these two genes can compensate null mutations in each other under nutrient-rich [yeast–peptone–dextrose (YPD)] conditions, but the cho2Δ deletant shows slow growth on glucose minimal medium (Fig. 4B). This indicates that, in the absence of exogenous choline, the Kennedy pathway (and hence PCT1), on its own, cannot support net phosphatidylcholine synthesis.

The second explanation for condition-dependence can be exemplified by the SAM1/SAM2 duplicate gene pair, which encodes two distinct forms of S-adenosylmethionine (AdoMet) synthetase. Although differentially regulated (24), the two genes can compensate null mutations in one another, and the double mutants are inviable under nutrient-rich (YPD) conditions (see Table 1). However, addition of AdoMet (the enzymatic product of Sam1p/Sam2p) into the medium yields viable double mutants (Fig. 4C).

Frequent Plasticity of Genetic Interactions Among Nonmetabolic Genes.

Having established the widespread occurrence of environment dependency of synthetic genetic interactions for metabolic genes, we asked whether condition dependency could be a general property of SSL interactions. First, we compiled a list of publicly available SSL interactions (25) discovered by a global genetic-interaction mapping approach (26) using a chemically defined glucose medium (those strains showing growth defects on minimal medium were excluded to ensure that the investigated interactions were not between unconditionally slow growing mutants, see Materials and Methods). Next, we collected viability data from published screens for single-gene-deletion phenotypes performed under 31 growth conditions (SI Table 2). In 57.4% of the investigated 2,666 SSL gene pairs, there is evidence that one or both members of the pairs make an essential contribution to growth under at least one of the 31 conditions investigated. This figure is likely to be an underestimate for two reasons: First, only a limited number of environments have been studied experimentally so far. Second, this estimate ignores cases where the double-deletant strain becomes viable under some other environmental condition. Moreover, there is some further support for a link between the extent of compensatory mechanisms and environmental specificity: genes for which evidence exists for conditional phenotypes have significantly more SL interactions than the rest of yeast's genes (Mann–Whitney U test, P = 0.002; see SI Fig. 6).

Random Phylogenetic Cooccurence of SSL Pairs.

Comparative genomics studies indicate that members of functional modules (i.e., genes that contribute jointly to a given cellular function) evolve nonindependently and show a similar phylogenetic distribution across species. For example, genes encoding members of protein complexes or metabolic modules are frequently gained and lost together during evolutionary history (27, 28). Indeed, we could confirm that gene pairs encoding subunits of the same literature-curated protein complexes (29) have higher phylogenetic cooccurrence than random gene pairs. We calculated a score (30) for the cooccurrence of these gene pairs across 16 eukaryotic genomes and used randomization protocols to get an estimate of statistical significance (see SI Materials and Methods). As expected, the score for subunits of protein complexes is significantly higher than expected by chance (P < 10⁻⁵, n = 7,186 pairs). Next, we asked if a similar result holds for experimentally determined SSL gene pairs. Using the same protocol as above, we found a strikingly different result. In contrast to members of protein complexes, gene pairs showing synthetic genetic interactions show no evidence for shared evolutionary history across species (P = 0.107, n = 1,850 pairs). Moreover, this finding cannot be explained by the likelihood of a low frequency of retention of redundant duplicates in all genomes; our result remains unchanged when all gene pairs showing even low sequence similarity to one another are excluded from the analysis (P = 0.113, n = 1,780 pairs, see SI Materials and Methods).

Analysis of Genetic Interactions in the Metabolic Network of Yeast.

We examined a recently updated (iLL672) metabolic network of S. cerevisiae, which contains 672 genes and 745 unique biochemical reactions including transport processes (18). The reconstruction also provides information on the association of genes with different metabolic subsystems (e.g., purine metabolism, phospholipid biosynthesis, etc.). One dubious reaction, corresponding to choline biosynthesis, was removed from the reconstruction because yeasts are unable to synthesize choline de novo (45). FBA of the metabolic network was used to calculate the impact of gene deletions on maximum biomass production rate (a proxy for fitness). Details of the FBA protocol have been described in ref. 12. SL interactions were identified by simulating all possible single- and double-gene deletions and screening for gene pairs where the single deletions had a <10% fitness effect, but the double mutant was unable to produce biomass (the use of different cut-offs led to very similar results). All deletion simulations were carried out in the ura3Δ leu2Δ his3Δ met17Δ lys2Δ genetic background to most closely mimic the strains used in the in vivo studies (see SI Materials and Methods).

To explore the condition dependency of SL interactions, we defined a large set of nutrient environments. First, we tested all external nutrients for their ability to support aerobic growth in minimal medium. This resulted in 50 minimal media containing different principal carbon sources, including glucose. Additionally, we defined a medium mimicking YPD, a medium where all possible external nutrients were allowed for uptake, and a minimal vitamin medium [lacking pantothenate because yeast is capable of de novo pantothenate biosynthesis (46)], resulting in 53 environmental conditions (for details see SI Materials and Methods and SI Table 8). All simulated growth media were supplemented with uracil, leucine, histidine, methionine, and lysine to complement the nutritional markers and also with vitamins (with or without pantothenate, see above) to further mimic the experimental conditions.

Experimental Procedures.

The simulations identified 59 gene pairs showing SL on either nutrient-rich (YPD) or glucose minimal [synthetic defined (SD)] medium. Published data (1, 18) on single-deletion phenotypes for these two conditions enabled us, in a comprehensive manner, to identify gene pairs for which the viability of single deletants was correctly predicted.

To carry out in vivo validation, we considered initially those gene pairs that have, at most, one paralog that is not annotated as an isozyme in the model. This choice enabled us to test higher-order genetic interactions by constructing triple-deletion strains in cases where the double mutant was viable (see below). Because, by using this criterion, all isozymes were excluded from validation, we additionally incorporated three randomly selected isozyme pairs in our experimental set. Moreover, among the group of gene pairs containing several paralogs, we decided to test gene pairs involved in pantothenate and polyamine biosynthesis for which we had the highest number of predicted SL interactions but no literature support available. Finally, some of the selected gene pairs could not be verified because one or the other mutant strain was missing from the deletion collection or contained a second-site mutation (18). This selection procedure left us with a set of 13 gene pairs to validate in vivo. These pairs corresponded to 17 cases of synthetic lethality: 6 on YPD, 3 on glucose minimal medium (SD), and 4 on both media.

We constructed the predicted double mutants by crossing haploid yeast strains containing single-gene deletions in the BY4742 and BY4741 backgrounds following standard protocols (see SI Materials and Methods).

In cases where we failed to detect any growth defect by visual inspection of plates (no overt synthetic sick or lethal phenotype), we performed accurate measurements of maximum growth rates of single and double mutants to estimate epistasis. Optical densities were measured by a Bioscreen C analyzer (Thermic Labsystems, Oy, Finland), and maximum growth rates were calculated by using an established protocol (19). Five cultures were grown for each strain in both YPD and SD. Maximum growth rates were averaged over the five replicates and divided by the wild-type value to yield a relative growth rate for each strain. Because additivity of the growth rates is equivalent to multiplicity of nonlogarithmic measures of fitness (47), we defined epistasis (ε) as the degree of departure from additivity of the relative growth rates (μ), thus ε = μ_AB + μ_ab − μ_Ab − μ_aB.

See SI Materials and Methods for details on the construction of triple-gene deletants.

Analysis of Global Genetic Interaction and Mutant Phenotype Data Sets.

We compiled a list of publicly available SSL interactions (25), discovered by the synthetic genetic array approach (26), on SD medium complemented with amino acids. Single-gene deletion strains that exhibit a pronounced growth defect (<80% of wild-type growth rate) on SD medium (48) were excluded from further analyses to ensure that the interactions were not between unconditionally slow growing mutants. This resulted in a list of 2,666 synthetic genetic interactions (1,230 of them being SL). Information on environment-specific phenotypes of single-gene deletions in nonessential genes was collected from published large-scale phenotypic screens (see SI Table 2). Our list of 31 experimental conditions included various nutrient and stress conditions, sporulation, and stationary phase, but excluded drug treatments. Only the strongest growth defects and phenotypes were considered as evidence for conditional fitness contributions.