Environmental robustness of the global yeast genetic interaction network

Environmental impacts on gene networks A phenotype can be affected by genes interacting with other genes, the environment, or both other genes and the environment (a differential interaction). To better understand how these interactions function in yeast, Costanzo et al. mapped gene-gene interactions using single- and double-mutant deletions and temperature-sensitive alleles under 14 environmental conditions. Many deleted or temperature-sensitive nonessential genes affected yeast fitness both positively and negatively under at least one of the environmental conditions tested. In these cases, up to 24% of yeast genes were affected. A minority of these differential interactions point to previously unknown genetic connections across functional networks, informing on how genetic architecture responds to environmental variation. Science, this issue p. eabf8424 The global yeast genetic interaction network is robust to environmental perturbation and elucidates the functional architecture of a eukaryote. INTRODUCTION Genetic interactions are identified when variants in different genes combine to generate an unusual phenotype compared with the expected combined effect of the corresponding individual variants. For example, a synthetic lethal genetic interaction occurs when two mutations, neither of which is lethal on their own, combine to generate a lethal double-mutant phenotype. Although there are millions of possible gene-gene combinations for any eukaryotic cell, only a rare subset of gene pairs will display a genetic interaction. Digenic, or gene-by-gene (GxG), interactions appear to underlie key aspects of biology, including the relationship between genotype and phenotype. Environmental conditions can modulate the phenotype associated with genetic variants, giving rise to gene-by-environment (GxE) interactions, when a single variant phenotype is modified, or gene-by-gene-by-environment (GxGxE) interactions when a genetic interaction is changed. RATIONALE A global genetic interaction network has been mapped for the budding yeast Saccharomyces cerevisiae, identifying thousands of connections that often occur between functionally related genes. Because the global genetic network was mapped in a specific reference condition, the potential for different environmental conditions to rewire the network remains unclear. Automated yeast genetics, combined with knowledge of a reference map, enables quantification of the extent to which new environmental conditions either modulate known genetic interactions or generate novel genetic interactions to influence the genetic landscape of a cell. RESULTS We tested ~4000 yeast single mutants for GxE interactions across 14 diverse environments, including an alternative carbon source, osmotic and genotoxic stress, and treatment with 11 bioactive compounds targeting distinct yeast bioprocesses. To quantify GxGxE interactions, we constructed ~30,000 different double mutants, involving genes annotated to all major yeast bioprocesses, and we scored them for genetic interactions. The plasticity of the network is revealed by differential genetic interactions, which occur when a genetic interaction observed in a particular condition deviates from that scored in the control reference network. Although ~10,000 differential interactions were discovered across all 14 conditions, we observed ~60% fewer differential interactions per condition as compared with genetic interactions in the reference condition, indicating that GxGxE interactions are rare relative to GxG interactions. On average, a single environmental perturbation modulated ~14% of the reference genetic interactions and revealed a smaller subset of ~7% novel differential interactions. Whereas GxG genetic interactions tend to connect pairs of genes that share a close functional relationship, novel differential GxGxE interactions mediate weaker connections between gene pairs with diverse roles. CONCLUSION Our general findings reveal how environmental conditions modulate the yeast global genetic interaction network, allowing us to assess the plasticity of genetic networks and the extent to which mapping genetic interactions in different environments can expand a reference network. Although different environments have the potential to reveal novel interactions and uncover previously unidentified but weaker functional connections between genes, the vast majority of genetic interactions remained unchanged in different conditions, suggesting that the global yeast genetic interaction network is largely robust to environmental perturbation. Systematic analysis of environmental impact on the global yeast genetic interaction network. (Top left) Mapping GxE interactions and (bottom left) GxGxE differential interactions reveals (top right) the environmental robustness of the global yeast genetic interaction network, (bottom right) highlighting new and distant functional connections associated with novel differential interactions. Phenotypes associated with genetic variants can be altered by interactions with other genetic variants (GxG), with the environment (GxE), or both (GxGxE). Yeast genetic interactions have been mapped on a global scale, but the environmental influence on the plasticity of genetic networks has not been examined systematically. To assess environmental rewiring of genetic networks, we examined 14 diverse conditions and scored 30,000 functionally representative yeast gene pairs for dynamic, differential interactions. Different conditions revealed novel differential interactions, which often uncovered functional connections between distantly related gene pairs. However, the majority of observed genetic interactions remained unchanged in different conditions, suggesting that the global yeast genetic interaction network is robust to environmental perturbation and captures the fundamental functional architecture of a eukaryotic cell.

INTRODUCTION: Genetic interactions are identified when variants in different genes combine to generate an unusual phenotype compared with the expected combined effect of the corresponding individual variants. For example, a synthetic lethal genetic interaction occurs when two mutations, neither of which is lethal on their own, combine to generate a lethal doublemutant phenotype. Although there are millions of possible gene-gene combinations for any eukaryotic cell, only a rare subset of gene pairs will display a genetic interaction. Digenic, or gene-by-gene (GxG), interactions appear to underlie key aspects of biology, including the relationship between genotype and phenotype.
Environmental conditions can modulate the phenotype associated with genetic variants, giving rise to gene-by-environment (GxE) interactions, when a single variant phenotype is modified, or gene-by-gene-by-environment (GxGxE) interactions when a genetic interaction is changed.
RATIONALE: A global genetic interaction network has been mapped for the budding yeast Saccharomyces cerevisiae, identifying thousands of connections that often occur between functionally related genes. Because the global genetic network was mapped in a specific reference condition, the potential for different environmental conditions to rewire the network remains unclear. Automated yeast genetics, combined with knowledge of a reference map, enables quantification of the extent to which new environmental conditions either modulate known genetic interactions or generate novel genetic interactions to influence the genetic landscape of a cell. RESULTS: We tested~4000 yeast single mutants for GxE interactions across 14 diverse environments, including an alternative carbon source, osmotic and genotoxic stress, and treatment with 11 bioactive compounds targeting distinct yeast bioprocesses. To quantify GxGxE interactions, we constructed~30,000 different double mutants, involving genes annotated to all major yeast bioprocesses, and we scored them for genetic interactions. The plasticity of the network is revealed by differential genetic interactions, which occur when a genetic interaction observed in a particular condition deviates from that scored in the control reference network. Although~10,000 differential interactions were discovered across all 14 conditions, we ob-served~60% fewer differential interactions per condition as compared with genetic interactions in the reference condition, indicating that GxGxE interactions are rare relative to GxG interactions. On average, a single environmental perturbation modulated~14% of the reference genetic interactions and revealed a smaller subset of~7% novel differential interactions. Whereas GxG genetic interactions tend to connect pairs of genes that share a close functional relationship, novel differential GxGxE interactions mediate weaker connections between gene pairs with diverse roles. CONCLUSION: Our general findings reveal how environmental conditions modulate the yeast global genetic interaction network, allowing us to assess the plasticity of genetic networks and the extent to which mapping genetic interactions in different environments can expand a reference network. Although different environments have the potential to reveal novel interactions and uncover previously unidentified but weaker functional connections between genes, the vast majority of genetic interactions remained unchanged in different conditions, suggesting that the global yeast genetic interaction network is largely robust to environmental perturbation. ▪ G enetic interactions are identified when two or more different gene variants combine to cause an unusual phenotype that deviates from a model on the basis of the combined effects of the corresponding single-variant phenotypes (1). Digenic or gene-by-gene (GxG) interactions appear to underlie diverse and fundamental aspects of biology, including the relationship between genotype and phenotype, the evolution of sexual reproduction, and speciation (2)(3)(4)(5)(6). The phenotype associated with genetic variants can also be modulated by environmental factorsincluding growth conditions, age, cell type, or microbe exposure-giving rise to gene-byenvironment (GxE) interactions or, if a genetic interaction is modified, gene-by-gene-byenvironment (GxGxE) interactions. Eukaryotic genomes contain thousands of different genes and millions of possible genetic interactions and function in many environmental contexts, meaning that our ability to explore the extent to which genetic interactions con-tribute to trait heritability is a major combinatorial and statistical challenge (3).
Efforts to address this challenge have involved the systematic mapping of genetic interactions in accessible model systems, primarily the budding yeast Saccharomyces cerevisiae. Large-scale application of an automated genetic approach, synthetic genetic array (SGA) analysis, enabled the majority of all possible gene pairs to be tested for negative and positive genetic interactions (7,8). A negative genetic interaction corresponds to a synthetic lethal or sick interaction when a double mutant shows a fitness defect greater than the expected effect of the combined single-mutant fitness phenotypes. Conversely, a positive genetic interaction is scored in a double mutant that grows better than expected (9). Genetic suppression represents an extreme type of positive interaction, in which the doublemutant fitness phenotype is greater than that of the least fit single mutant (10). The resultant global digenic interaction network comprises 350,000 positive and~550,000 negative interactions, which tend to connect pairs of functionally related genes (11,12). Genes encoding members of the same biological pathway or protein complex often share similar patterns or profiles of negative and positive genetic interactions, and a global network clusters genes with similar genetic interaction profiles together into a hierarchy of organized modules, corresponding to protein complexes or pathways, biological processes, and cellular compartments, revealing the functional architecture of a eukaryotic cell (Fig. 1A) (11,12).
The reference yeast genetic network was mapped in a single genetic background in a specific reference condition. The phenotypes of both single-and double-mutant cells can be modulated by the environment, which leads to GxE and more complex GxGxE interactions. For example, subsets of functionally biased yeast genes have been surveyed in response to several different environmental conditions, and condition-specific genetic interactions have been identified (13)(14)(15)(16)(17)(18). However, given the focus so far on specific bioprocesses, such as DNA repair, the plasticity of genetic networks across a broad array of environmental conditions remains unclear. We systematically assessed GxE and GxGxE effects in the context of the structure and topology of the global yeast genetic interaction network. This powerful model system allows us to assess the extent to which the environment modulates genetic interactions to influence the overall genetic landscape of a cell and provides insights into the utility of genetic network reference maps.

Single-mutant fitness and gene-environment interactions
Quantitative analysis of digenic interactions depends on the accurate assessment of singleand double-mutant phenotypes (19). Therefore, to accurately measure the effect of environment on the yeast genetic network, a baseline of quantitative measurements of conditionspecific effects on single-mutant fitness (GxE interactions) is required. To identify GxE interactions in the context of a SGA screen, we subjected an array of single mutants to SGA analysis using a neutral (control) "query" locus and assessed colony size after growth in different conditions. We examined 14 diverse conditions, including an alternative carbon source, osmotic stress, genotoxic stress, and 11 bioactive compounds that target distinct yeast biological processes within the global genetic interaction profile similarity network (Fig. 1, A and B). Relative fitness measurements were obtained for 3704 viable deletion mutant strains and 782 temperature-sensitive (TS) alleles, corresponding to 553 essential genes, both in the reference SGA condition and in each of the 14 diverse conditions (data file 1) (20)(21)(22)(23).
To identify GxE interactions from our singlemutant fitness measurements, we calculated a differential fitness score for each mutant by measuring the difference between mutant fitness in the reference condition versus each of the 14 conditions (fig. S1A) (23). More than half of all the genes that we examined (59%) were associated with a significant differential single-mutant fitness defect (P < 0.05, differential fitness score < -0.08) (Methods summary) (23) in at least one condition (fig. S1B). The union of all genes that displayed fitness defects in each condition revealed that six to eight conditions identified the majority of distinct genes associated with condition-specific growth defects (fig. S1C). Additional conditions captured relatively few new genes, perhaps because different conditions can lead to a similar stress response (24). Thus, our set of 14 conditions affects a wide range of functionally diverse genes and should be useful for exploring how different environments affect overall structure and topology of the global genetic interaction network.
The total number of strains with detectable differential fitness defects in a single condition ranged from~250 (4.5%) to~1200 (22%), with the largest number of condition-specific fitness defects observed in the presence of Benomyl, a microtubule depolymerizing agent (fig. S1D). Stronger negative differential fitness scores were associated with genes encoding known targets of compounds ( fig. S1A). More generally, genes with differential fitness defects in a particular growth condition were often involved in a specific function perturbed by the corresponding growth condition ( fig. S2 and data file 2) (20). For example, in Benomyl, numerous genes involved in mitosis or DNA replication and repair were enriched for differential fitness defects ( Fig. 1C and data file 2) (20). A number of genes involved in mRNA processing also displayed fitness defects in Benomyl, which may reflect that genes encoding α-tubulin, TUB1, and TUB3 are among the relatively few yeast genes with introns ( Fig. 1C and data file 2) (20,25). By contrast, the differential fitness defects in Monensin, an intracellular traffic inhibitor (26), were most enriched among genes involved in vesicle trafficking, glycosylation, and cell wall biosynthesis ( Fig. 1D and data file 2) (20).
Some mutants (1083) had positive differential fitness scores, indicating better growth relative to a wild-type strain in a particular condition compared with the corresponding mutant growth relative to wild-type in the reference condition ( fig. S1, A, B, and D). For example, although all strains, including wild type, grew more slowly in the presence of the proteasome inhibitor Bortezomib, strains carrying TS alleles of essential genes were enriched among the set of mutant strains that showed positive differential fitness scores in this condition (P < 10 −27 , Fisher's exact test) ( fig. S3A). This finding mirrors a previous Costanzo (11). The similarity network was annotated by using SAFE (54), identifying network regions enriched for similar GO biological process terms, which are color-coded. (B) Conditions and query genes selected for this study. Location of each query gene on the global genetic interaction profile similarity network is indicated. Bioprocesses targeted by selected bioactive compounds are also shown. (C) Regions of the global similarity network significantly enriched for genes exhibiting differential single-mutant fitness defects in Benomyl. (D) Regions of the global similarity network significantly enriched for genes exhibiting differential single-mutant fitness defects in Monensin. For both (C) and (D), regions of the global similarity network significantly enriched for genes exhibiting negative differential fitness defects were mapped by using SAFE. Examples of genes located in the most enriched regions are indicated in blue. observation from analysis of the global genetic interaction network in which hypomorphic alleles of proteasome genes show positive genetic interactions, which likely reflects reduced degradation of the products of TS alleles, providing a growth advantage (11). Similarly, approximately sixfold more TS alleles showed positive differential fitness, relative to those with a negative differential fitness, when grown in sorbitol, which may reflect a general effect of increased intracellular osmolarity (P < 10 −13 , Fisher's exact test) ( fig. S3B) (27). By contrast, TS alleles of essential genes were depleted for positive differential fitness effects in certain conditions, including growth in the presence of the HSP90 inhibitor Geldanamycin (P < 10 −4 , Fisher's exact test) ( fig. S3A). This suggests that the fitness of different TS mutants is dependent on HSP90 activity, which is consistent with the known role of HSP90 as a phenotypic capacitor (11,(28)(29)(30).
Last, we compiled a list of~350 dynamic, "condition-responsive" genes, corresponding to the top~5% of all tested genes based on the number of conditions in which they showed either a significant positive differential fitness score in two or more conditions or a significant negative differential fitness score in four or more conditions (data file 1) (20). Consistent with our findings, conditionresponsive genes identified in our assay were also often sensitive to multiple chemicals when assayed as homozygous gene deletion mutants (2.7-fold, P < 10 −12 , Fisher's exact test) ( fig. S3C) (31). Moreover, condition-responsive genes were more likely to be essential or to have a significant growth phenotype as a deletion mutant in the reference condition, were often highly connected on the global yeast genetic interaction network, and often encoded highly conserved, multifunctional proteins ( fig. S3D).

Mapping genetic interactions across different environments
Our systematic assessment of GxE interactions involving single mutants provides a foundation for comprehensive analysis of how genetic interactions are modulated by environment (GxGxE interactions). To do this, we took advantage of a diagnostic array used previously to survey complex, trigenic interaction networks (data file 1) (6,20). This array consists of 1200 strains, comprising~1000 nonessential gene deletion mutants and~200 TS alleles of essential genes, spanning~20% of all yeast genes and representing the functional breadth of the yeast genome (data file 1) (6,20). Mutant strains included on the diagnostic array exhibited a range of fitness defects and are generally representative of the genome-wide distribution of differential single-mutant fitness effects (data file 1) (20). For conditionspecific genetic interaction analysis, we selected 26 query genes, each of which shows a substantial number of genetic interactions in the global network (11) and a functionally coherent genetic interaction profile that localizes the query gene to a specific biological process cluster on the reference genetic interaction similarity network (Fig. 1B).
We crossed the 26 query strains to the diagnostic mutant array, resulting in~30,000 doublemutant strains, which were each assessed for environmental modulation of genetic interactions in all 14 conditions. To measure condition-dependent genetic interactions, every double-mutant array generated from a single-query gene was screened three times. One copy was grown in the standard SGA condition, whereas the two other copies were each grown in different conditional media ( fig.  S4). This configuration provided a matched reference control for every condition, which facilitated normalization of systematic experimental artifacts and improved accuracy of condition-specific genetic interaction measurements (32). The entire screening pipeline was repeated twice, providing two independent biological replicates for every query gene in every test condition, each one with four separate double-mutant colonies (eight tests of each double mutant per condition) and 14 independent biological replicates of each query screen in the standard reference SGA condition (23).
Negative and positive genetic interactions were quantified as previously described (19). We identified an average of~2100 negative and~1400 positive genetic interactions on the basis of 14 biological replicate screens performed in the standard reference condition, at an intermediate confidence threshold (fig. S5A and data file 3) (20,23). A similar number of gene pairs exhibited a genetic interaction in any single test condition, with an average of~2150 negative and~1500 positive genetic interactions identified per condition ( fig. S5A). Our genetic interaction measurements were reproducible, and those interactions identified in the reference condition in this study overlapped substantially with interactions derived from our previous genomewide study ( fig. S5B) (11). Comparison of the overlap of genetic interaction profiles between a specific condition screen and its matched control with that seen with an unmatched control suggested that comparison with a matched reference provides the necessary sensitivity to detect rare condition-specific interactions ( fig. S5C, tables S1 and S2, and data file 3) (20), which is consistent with previous studies (32).

Quantifying and classifying differential interactions
To discover differential or GxGxE interactions in our scored data, we quantified the differ-ence between genetic interaction scores derived from each conditional screen and its corresponding matched reference ( fig. S4) (23). We scored a differential negative interaction when a genetic interaction score observed in a particular condition was less than the genetic interaction score measured in the corresponding reference control. Conversely, a genetic interaction that was stronger in a given condition as compared with the matched reference control was scored as a differential positive interaction. Analysis of biological replicates confirmed the reproducibility of differential interaction scores ( fig. S5D and tables S1 and S2) (23).
The Benomyl screen showed the largest number of differential, condition-specific fitness defects and genetic interactions, providing a rich context for developing a general scheme for defining various classes of differential interactions. We measured Benomyl differential interactions as described above, adopting an intermediate confidence threshold ( D j j > 0:08, P < 0.05) (Methods summary and Fig. 2A, left) (23). In total, we identified 1367 differential interactions in Benomyl, which was~60% fewer than the 3845 genetic interactions identified in the matched reference condition (Fig. 2B). Thus, the majority of the genetic interactions observed in the reference condition were also observed in the presence of Benomyl. To determine whether the prevalence of Benomyl differential interactions derived from the diagnostic array was generalizable to the whole genome, we also screened the same set of 26 SGA query mutant strains against the complete collection of nonessential gene deletion mutants and essential gene TS alleles, in the absence or presence of Benomyl ( fig. S6, A and B, and data file 3) (20,23). This analysis also revealed~65% fewer differential interactions relative to genetic interactions, indicating that trends observed with the diagnostic array accurately reflect an unbiased genome-scale analysis ( fig. S6C).
In an effort to discover what new functional information might be associated with differential interactions, we divided them into four classes: "reversed," "modified," "masked," and "novel" (data file 3) (20, 23). These classes of differential interactions can be depicted schematically for those with either negative ( Fig. 2A, right) or positive ( fig. S7A) scores. The reversed class was the rarest and included gene pairs that showed significant but opposite genetic interactions in a condition versus the matched reference, accounting for less than 1% of all Benomyl differential interactions (Fig. 2B). Often these differentials involved at least one relatively weak interaction, which means that they are likely more prone to false positives and false negatives. Thus, the sign of a fitness-based genetic interaction does not frequently change in an altered environment. The remaining classes were roughly equivalent in size and accounted for the majority (>99%) of all Benomyl differential interactions ( Fig. 2B and data file 3) (20).
We identified 538 (538 of 1367;~39%) novel differential interactions, which were not observed as genetic interactions in the reference network. The remaining (829 of 1367;~61%) differential interactions were either modified or masked, meaning that they were scored as genetic interactions in a particular condition as well as in the reference control but differed (B) The number of genetic and differential interactions identified from SGA screens performed in Benomyl and in the reference condition. The fraction of significant reference condition genetic interactions that do not exhibit a significant differential interaction are shown in black. The fraction of differential interactions is shown in gray and white. (Inset) The outer black and gray rings indicate the size of the reference and differential interaction networks mapped in the reference and Benomyl conditions, respectively. The larger chart summarizes the total number of negative and positive Benomyl differential genetic interactions identified at an intermediate score threshold. The fraction of negative and positive interactions classified as reversed, novel, modified or masked is indicated. Specific classes of differential negative interactions are indicated and colored in shades of blue. Specific classes of differential positive interactions are indicated and colored in shades of yellow. Examples of gene pairs in each interaction subclass along with their corresponding reference, condition, and differential genetic interaction scores are shown.
in relative strength, and thus were modulated by the environment. Genome-wide analysis revealed similar fractions of modified, masked, and novel Benomyl differential interactions ( fig. S6D). Analysis of specific examples of modified differential interactions from the Benomyl screens revealed that modified differential interactions could arise in two ways ( Fig. 2A  and fig. S7A). First, a genetic interaction detected in the reference network may be exacerbated in a specific condition, resulting in a significant differential score. For example, a negative genetic interaction between MYO2 and DYN2 was stronger in Benomyl, reflecting the role of these two motor proteins in nuclear positioning and spindle orientation, especially when microtubule function is compromised (Fig. 2B). In another example, a positive genetic interaction between the proteasome gene RPN12 and TCP1, which encodes an essential subunit of the chaperonin-containing T-complex involved in tubulin folding (Fig. 2B), was enhanced in the presence of Benomyl. In this case, compromising proteasome function may impair the degradation of the TCP1 TS allele product, which is particularly relevant in the presence of Benomyl.
Second, a negative differential score may arise when a positive genetic interaction in the reference network is weakened in a specific condition ( Fig. 2A and fig. S7A). For example, a positive genetic interaction between GIM3 and RPN12, which encode subunits of the Prefoldin chaperone complex and the 19S proteasome, respectively, was weaker in the presence of Benomyl, resulting in a negative differential score (Fig. 2B). Similarly, a positive differential score can result from a conditionspecific change in magnitude of a negative genetic interaction. For example, in the reference condition, GIM3, which is involved in tubulin folding, showed a relatively strong negative genetic interaction with DAD1, which is an essential kinetochore gene (Fig. 2B). In Benomyl, the GIM3-DAD1 negative genetic interaction was significant but much weaker, resulting in a differential positive interaction (Fig. 2B), which may reflect that Benomyl perturbs the microtubule cytoskeleton in a manner that encompasses the cell physiology associated with the GIM3-DAD1 interaction.
We also categorized a specific subset of modified differential interactions, which we call masked, in which positive or negative genetic interactions were only identified in the reference condition ( Fig. 2A and fig. S7A). For example, whereas a negative genetic interaction was scored in the reference condition for GIM3 and TUB3, which encodes α-tubulin, no genetic interaction was observed in the presence of Benomyl, leading to a GIM3-TUB3 differential positive interaction (Fig. 2B). Positive genetic interactions scored in the reference con-dition can also be masked. For example, a positive genetic interaction between the yeast actin gene ACT1 and the microtubule motor encoding gene KAR3 was no longer detectable in Benomyl, leading to an ACT1-KAR3 differential negative interaction (Fig. 2B).
Although relatively rare, novel differential interactions are noteworthy because they are revealed only in a specific condition, but not in the reference condition, and thus should reflect functional links between genes that are driven by condition-dependent cellular physiology ( Fig. 2A and fig. S7A). For example, we identified a novel negative differential interaction between DSN1 and KIP3 in Benomyl that highlights a functional link between the kinetochore and a microtubule motor protein involved in spindle assembly (Fig. 2B).

Multiple environments, differential interactions, and the reference genetic network
We applied our scoring and classification system for differential genetic interactions, described above, to all 14 conditions surveyed. On average, less than 3% of all gene pairs tested in any given condition showed a differential interaction compared with~13% of all gene pairs that exhibited a genetic interaction in the reference condition, which is consistent with the percent of overlapping gene pairs that show a genetic interaction in the context of a genome-wide study ( fig. S7, B and C, and data file 3) (11,20). Comparison of negative interactions, in particular, revealed a notable difference in which~1% of tested gene pairs exhibited a significant differential negative interaction, relative to~8% of all gene pairs that showed a negative genetic interaction in any single condition tested (Fig. 3A, fig. S7C, and data file 3) (20). Two-thirds of all differential interactions were classified as modified or masked because they overlapped a genetic interaction in the reference condition (Fig. 3B). On average, modified and masked differential interactions from a single condition accounted for~14% of all genetic interactions in the reference network (Fig. 3B). Depending on the condition, we estimate that between~5 and 24% of genetic interactions detected in a reference genetic network can be modulated in a different environment.
Novel differential interactions provide a direct estimate of the additional functional information that different environments can contribute to a genetic network. On average, when compared with a matched reference control,~7% of genetic interactions identified in a single condition were classified as novel differential interactions ( Fig. 3B and data file 3) (20). Thus, the vast majority (~93%) of the genetic interactions mapped in different environmental conditions were also on the reference network. Novel differential negative interactions were also significantly weaker in magnitude as compared with genetic interactions measured in the reference condition (Fig. 3C). Detailed analysis of previously reported differential interaction networks (14,15,17) confirmed that differential interactions-in particular, novel differential interactionswere much less abundant than genetic interactions ( fig. S7D).
It is critical to account for the false negative rate associated with high-throughput genetic interaction screens because failure to detect a true reference condition genetic interaction can be mistakenly classified as a novel differential interaction. We applied a Markov chain Monte Carlo (MCMC) modeling approach to generate a robust consensus set of negative and positive reference condition genetic interactions for each of the 26 SGA query genes based on the collection of 14 independent, biological replicate screens (23). The resultant consensus genetic interaction profiles were used as a gold standard to estimate false discovery and false negative rates at defined confidence thresholds (tables S1 and S2 and data file 4) (20). From analysis of reference condition replicate screens, we estimated false negative rates of 39 and 52% for negative and positive genetic interactions, respectively (table S1) (23). In the case of Benomyl, we mapped 538 novel differential interactions when compared with a reference condition network derived from a single, matched control ( Fig. 3D  and fig. S7C). However, a more rigorous comparison of Benomyl genetic interactions versus the MCMC consensus reference genetic interaction profile identified 454 novel differential interactions or~15% fewer interactions in comparison with a matched control screen (Fig. 3D). A third comparison using a saturated reference network-on the basis of a reference network derived from the union of all the control screens ( fig. S7E), which is less prone to false-negative interactions-identified 319 novel differential interactions, or~40% fewer interactions as compared with those identified by using a single matched reference control (Fig. 3D) (23). The number of novel differential interactions could decrease by as much as~60% depending on the condition and whether the reference was based on a single matched control, a consensus network control (23), or the union of reference control screens ( Fig. 3D and data file 4) (20). Our analysis highlights the importance of a robust reference network, replicate screens, and rigorous estimates of false-negative rates for comparative study of genetic interactions in an alternative environmental condition.

Properties of differential interactions
Comparing all 14 conditions with their matched controls identified a combined total of~10,000 differential interactions, the majority (61%) of Costanzo Fig. 3. The relative contribution of reference genetic interactions and differential interactions to the yeast genetic network. (A) Box plots showing the distribution of negative genetic and differential negative interaction density (total number of interactions per total gene pairs tested) per condition and per query mutant screened. The dotted line indicates the average genetic interaction density for the same set of array genes in the global genetic interaction network (11). (B) The average fraction of genetic and differential interactions identified from SGA screens performed in the reference and one additional condition. The fraction of reference condition genetic interactions that do not exhibit a significant differential interaction are shown in black. Modified and masked differential interactions, which overlap with interactions identified in the reference condition, and novel differential interactions are indicated. The outer black and gray rings indicate the average size of the reference and differential interaction networks mapped for one additional condition, respectively. The colored diagram summarizes the total number of negative and positive differential genetic interactions identified at an intermediate score threshold from analysis of 14 different test conditions. The fraction of negative (shades of blue) and positive (shades of yellow) interactions classified as reversed, novel, modified, or masked is indicated. (C) Distribution of negative genetic interaction scores (light blue) and novel differential negative interaction scores (dark blue). (D) The number of novel differential interactions identified per condition by using different reference genetic interaction networks. which were specific to a single growth condition, with differential negative interactions being less prevalent and relatively weaker than differential positive interactions (Fig. 3B  and fig. S7F). Genes that were highly connected hubs on the global genetic network were also more likely to be hubs on a differential network because the average number of differential interactions for an individual array gene, across all 14 conditions, was significantly correlated to the interaction degree in the global genetic network [Pearson correlation, correlation coefficient (r) =~0.7, P < 10 −16 ] (fig. S8, A and B). The number of novel differential interactions associated with each query mutant examined in this study was also cor-related to interaction degree in the global genetic network (Pearson correlation, r =~0.5, P < 0.006) ( fig. S8C), suggesting that the frequency of novel differential interactions observed by using a diagnostic set of genes should reflect the genome-wide prevalence of novel differential interactions. Consistent with these observations, gene features associated with hubs on the reference network were shared with differential network hubs ( fig. S8D) (11). Notably, high novel differential interaction degree was associated with genes whose loss of function resulted in a single-mutant fitness defect in the reference condition ( fig. S8E), and genes with condition-dependent fitness defects had proportionally more novel differ-ential interactions than those of genes lacking condition-dependent fitness defects ( fig. S8F). Like genetic interaction hubs, genes with many novel differential interactions were associated with multiple Gene Ontology (GO) annotations and lower dN/dS [the ratio of the number of nonsynonymous substitutions per nonsynonymous site (dN) to the number of synonymous substitutions per synonymous site (dS)], suggesting that they are more functionally pleiotropic and tend to be under stronger evolutionary constraints ( fig. S8, D  and E). Conversely, transcript levels of high differential interaction degree genes did not vary substantially across different environments or genetic backgrounds, suggesting that Costanzo Fig. 4. Functional evaluation of differential interactions. (A) Plots of precision versus recall for negative genetic and differential interactions, as determined by our genetic interaction score ( D j j > 0:08, P < 0.05). True-positive (TP) interactions were defined as those involving gene pairs co-annotated to a gold standard set of GO terms, as defined elsewhere (55). False-positive (FP) interactions were defined as those involving pairs of genes annotated to different GO terms, as defined elsewhere (55). The background precision at which true positives are randomly identified is indicated by the dotted line. The precision and recall values were calculated as previously described (19). (B) Fold enrichment for negative (blue) and positive (yellow) genetic and differential interactions among colocalized, coexpressed, physically interacting, or co-complexed gene pairs or their encoded proteins were calculated for genetic interactions identified in the reference condition, 14 conditions, and each differential interaction class. Comparisons based on fewer than 10 overlapping gene pairs are indicated (open boxes). (C and D) Plots of precision versus recall (number of TP) for positive genetic and differential interactions, as determined by our genetic interaction score ( D j j > 0:08, P < 0.05). True-positive interactions were defined as those involving gene pairs co-annotated to a gold standard set of GO terms, as defined elsewhere (55). False-positive (FP) interactions were defined as those involving pairs of genes annotated to different GO terms, as defined elsewhere (55). The background precision at which true positives are randomly identified is indicated with the dotted line. The precision and recall values were calculated as described (19). environmentally responsive gene expression patterns are not generally predictive of differential interactions ( fig. S8D).
Negative genetic interactions measured in any condition tended to connect functionally related gene pairs and overlapped with other types of molecular interaction networks (Fig. 4, A and B). Differential negative interactions also identified functionally related gene pairs, but to a lesser extent (Fig. 4,  A and B). Closer examination revealed that most of the functional signals associated with differential negative interactions were captured by the modified class, whereas those belonging to the novel or masked categories did not overlap substantially with other molecular interaction datasets (Fig. 4B). Thus, negative genetic interactions that occurred in a condition-specific manner (novel differential negative interaction) ( Fig. 2A) and positive genetic interactions that are masked in a particular condition (masked differential negative interaction) ( Fig. 2A) often involve gene pairs with unrelated functional annotations.
Consistent with previous observations (11), positive genetic interactions identified in either the reference or alternative conditions were less functionally informative than negative genetic interactions (Fig. 4, A and C). However, the complete set of differential positive interactions appeared to connect functionally related genes more often than positive genetic interactions measured in either the reference or individual test conditions (Fig. 4C). Modified and masked differential positive interactions were the most predictive of functionally related gene pairs, often connecting members of the same protein complex and overlapping substantially with protein-protein interactions (Fig. 4B). Further analysis revealed that the functional signal associated with the masked and modified classes was largely attributable to gene pairs that displayed negative genetic interactions in the reference condition but were either weaker (modified) or no longer detectable (masked) in a particular condition (Fig. 4D). This may reflect a particular environmental perturbation that mimics the cellular physiology associated with a double-mutant strain grown in the reference condition, obscuring a phenotype associated with a genetic interaction. Thus, although modified and masked differential positive interactions tend to connect functionally related gene pairs, this same information is captured by negative genetic interactions identified in the reference condition. Genome-wide analysis revealed similar functional trends associated with Benomyl differential negative and positive interactions ( fig. S6, E and F).
To further explore the functional information associated with condition-specific interactions, we grouped together array genes that belong to the same biological process cluster represented on the global genetic profile similarity network (Fig. 1A) and measured how often each query gene showed either genetic or novel differential interactions with each functional group (Fig. 5 and data file 5) (11,20,23). Query and array genes within the same bioprocess cluster were often connected by negative genetic interactions in the reference condition (~3.6-fold enrichment within bioprocess) (Fig. 5A, on diagonal). For example, in the reference condition, the VTI1 query gene, which encodes an essential SNAP [soluble N-ethylmaleimide-sensitive factor (NSF) attachment protein] receptor (v-SNARE) that is involved in multiple protein sorting pathways (33)(34)(35) and located in the vesicle traffic bioprocess cluster, showed strong enrichment for negative genetic interactions with functionally related array genes located in the same vesicle traffic bioprocess cluster (Fig. 5B).
By contrast, novel differential negative interactions did not connect query and array genes annotated to the same biological process (Fig.  5C). Although VTI1 showed the most novel differential interactions of any query gene tested (data file 3) (20), most of these interactions did not involve other vesicle trafficrelated genes. Instead, we observed modest but significant enrichment for novel differential negative interactions between the VTI1 query gene and array genes with roles in cell polarity and nuclear transport (Fig. 5D). The vast majority (1489 of 1553,~96%) of all novel differential negative interactions identified by using a matched reference control connected pairs of genes located in different bioprocess clusters on the global similarity network. These results further indicated that condition-specific genetic interactions do not identify gene pairs with a close functional relationship in the same general bioprocess but rather have the potential to uncover weaker functional associations between distinct biological processes.

Differential interactions capture distant but coherent functional relationships
We next explored the functional distribution of novel differential interactions across each of the 14 different conditions. In particular, we tested whether array genes annotated to the same function showed more novel differential interactions in a particular condition (Fig. 6). Array genes annotated to specific biological processes were enriched for novel differential negative interactions in response to a specific condition that perturbs the same bioprocess (~2.3 fold enrichment within bioprocess) (Fig. 6A, on-diagonal). For example, consistent with previous observations (14, 15), we found that array genes in the DNA replication and repair bioprocess cluster were enriched for novel differential negative interactions in the presence of methyl methanesulfonate (MMS) (3.9-fold) (Fig. 6B). But as shown above (Fig. 5), these novel differential interactions did not involve related query genes with roles in DNA replication and repair. The enrichment observed among DNA replication and repair array genes was predominantly driven by MMS-specific novel differential negative interactions with the MYO2 query gene, a type V myosin motor involved in actin-based vesicle transport and spindle orientation, and with the VTI1 and TRS20 query genes, which are involved in vesicle transport, highlighting a functional link between DNA replication and vesicle trafficking (Fig. 6B) (33)(34)(35). In another example, vesicle traffic genes were enriched for novel differential negative interaction in response to Monensin (2.5-fold) (Fig. 6A). In this case, Monensin-specific negative interactions connected vesicle traffic array genes to the RSP5 query gene, which encodes a ubiquitin ligase involved in multivesicular body sorting, the heat shock response, endocytosis, and ribosome assembly (36), as well as the LSM6 query gene, which has a general role in RNA processing (Fig. 6B) (37,38). Novel differential positive interactions were also not enriched among functionally related genes, and they were less informative of gene function than were novel differential negative interactions ( fig. S9A and data file 5) (20).
A similar trend was observed in genomescale Benomyl screens, in which novel differential negative interactions specifically connected array genes with mitosis-related roles to functionally diverse query genes, such as NUP188, which encodes a nuclear pore component (39), and GPI15, involved in glycosylphosphatidylinositol (GPI) anchor biosynthesis ( fig. S9B and data file 5) (20,40,41). Thus, environment may sensitize genes with roles in a specific bioprocess to negative genetic interactions with functionally diverse query genes.
The majority of differential interactions overlap a genetic interaction identified in the reference condition, which often connect functionally related genes within the same biological process. The environmental rewiring of genetic networks is driven by rare conditionspecific and relatively weak novel differential interactions, which identify new connections between genes with diverse functions in different biological processes (Fig. 7).

Discussion
We surveyed a set of functionally diverse yeast genes and different environments, quantifying how the growth phenotypes associated with different single-gene mutations were modulated by the environment (GxE) and how the growth phenotypes associated with different genetic interactions (GxG) respond dynamically to a particular condition to generate differential interactions (GxGxE). Our general findings reveal how environmental conditions modulate the yeast global genetic interaction Costanzo ) in the reference condition were tested for enrichment for array genes in each of the biological processes indicated (y axis). Node size reflects the statistical significance of enrichment, and the shade boxes along the diagonal indicate instances in which the query and array genes belong to the same biological process cluster on the global genetic interaction profile similarity network (11). Dotted lines indicate bioprocesses enriched for negative genetic interactions with the VTI1 query gene. The average fold enrichment of negative genetic interactions within and between specific biological processes network, allowing us to assess the plasticity of genetic networks and the extent to which mapping genetic interactions in different environments can expand a reference network and generate new functional information.
In general, genetic interactions, especially negative genetic interactions, are rich in functional information because they tend to connect genes that function within the same biological process (Fig. 7B) (11). Analogously, if an environmental perturbation affects a particular biological process, then genes with roles in the perturbed bioprocess tend to show differential sensitivity in the corresponding condition (Fig. 7A). Most genetic interactions are not modulated by the environment and remain detectable in both a given test condition and the reference control condition. However, subsets of genetic interactions can be modified by a particular condition, leading to a differential interaction (Fig. 7, C and D). Differential interactions are relatively rare because, on average, we observed approximately threefold more genetic interactions in the reference condition as compared with differential interactions detected in any other single environment.
Most differential interactions belonged to the modified or masked classes, which overlapped with a negative or positive genetic interaction in the reference genetic network. The majority (~70%) overlapped specifically with negative genetic interactions and thus were functionally informative, connecting genes belonging to the same general biological process (Fig. 7C). However, because they recapitulate connections that are captured in the global genetic network mapped in the reference condition, modified and masked differential interactions do not contribute new functional information. On the other hand, novel differential interactions correspond to condition-specific genetic interactions between genes that do not interact in the reference condition. The rare subclass of novel differential negative interactions does not typically include functionally related gene pairs. Instead, these interactions tend to connect groups of genes that are sensitive to a particular environmental perturbation to functionally distant genes (Fig. 7D). Thus, novel differential interactions highlight new connections between distinct cellular functions and, as a result, have the potential to expand the global genetic network.
Although detecting novel differential interactions promises to add new information to genetic networks, an accurate definition of novel differential interactions is challenging because it depends on the quality and comprehensiveness of the reference genetic network. For example, multiple independent replicate screens reduce the number of falsenegative interactions observed in the reference condition but consequently also decrease the number of differential interactions classified as novel upon environmental perturbation (Fig. 3D). The relative fraction of novel differential interactions contributed to the reference network by a single environmental condition can vary widely when using a single, matched control (~7%) or the union of all available replicate controls (~1%) (data file 4) (20). Using a high-confidence consensus reference network (data file 4) (20), we estimate that screening one additional environmental condition contributes less than~4% (191 of 5174) novel interactions relative to the reference network, suggesting that most genetic interactions are captured in a single condition. Hence, although differential interaction analysis in multiple diverse conditions reveals the subset of genetic interactions that are modulated upon environmental perturbations (modified or masked interactions), they can only marginally expand the size of the network (novel interactions), highlighting that the global genetic interaction network is generally robust to environmental perturbations.
Although our study involved the use of a diagnostic array of yeast genes, selected environmental conditions, and functionally diverse query genes, several observations suggest that our results capture the general resilience of the global yeast genetic interaction network to environmental perturbation. First, our genomescale single-mutant fitness analyses suggested that the selected set of conditions and small molecules elicited widespread but distinct cell physiological effects. Second, a genome-wide comparison of genetic interactions measured in the absence and presence of Benomyl yielded a similar fraction of differential interactions seen by using a diagnostic mini-array of genes. Last, analysis of several independent interaction datasets-derived from SGA-based approaches by using various subsets of yeast genes and a number of different conditions, including various DNA damaging agents and stress-response conditions-confirmed that differential interactions are substantially less prevalent than the genetic interactions observed in the reference condition ( fig. S7D) (14,15,17).
Consistent with our results, previous surveys of subsets of genes involved in a specific cellular function perturbed by a particular condition revealed an enrichment for differential interactions that often occurred between functionally distant gene pairs (14,15,17). Although included in previous studies, a subset of differential interactions, based on statistically insignificant genetic interaction scores, was omitted from our analysis (23). However, these particular differential interactions were not enriched among functionally related genes and did not appreciably increase the total number of differential interactions relative to the size of the reference genetic interaction network examined here or reported in other studies (figs. S7D and S10). Thus, in general, genetic interactions between the vast majority of genes and their corresponding functional modules (such as complexes and pathways) are not dynamic or rewired in response to environmental stimuli. Moreover, we found that the~20% of yeast genes with relatively sparse genetic interaction profiles (11) are statistically depleted for condition-specific fitness defects (P < 10 −99 , Fisher's exact test, one-tailed). The relationship between fitness and interaction degree ( fig. S8) along with the strong correlation observed between interaction degree for a given gene in the global genetic and differential networks ( fig. S8) further suggests that condition-specific genetic interactions will not appreciably increase the number of interactions associated with low-degree genes in the global reference genetic network.
Differential interactions reflect the phenotypic consequences of combining three independent perturbations, two genetic and one is shown in the box plot. (B) Regions of the global genetic interaction profile similarity network significantly enriched for array genes exhibiting negative genetic interactions with the VTI1 query gene in the reference condition were mapped by using SAFE (54). The functional regions of the global similarity network significantly enriched for interactions with VTI1 are indicated with blue dotted lines. Array genes enriched for negative genetic interactions are shown in blue. The location of the VTI1 query gene on the global similarity network is indicated by a white node. (C) Novel differential negative interactions for each query gene (x axis) across all 14 test conditions were tested for enrichment for array genes in each of the biological processes (y axis). Node size reflects the statistical significance of enrichment, and the shaded boxes along the diagonal indicate instances in which the query and array genes would have belonged to the same biological process cluster on the global genetic interaction profile similarity network (11). Dotted lines indicate bioprocesses enriched for novel differential negative interactions with the VTI1 query gene. The average fold enrichment of novel differential negative interactions within and between specific biological processes is shown in the box plot. (D) Regions of the global genetic interaction profile similarity network significantly enriched for array genes exhibiting novel differential negative interactions with the VTI1 query gene were mapped by using SAFE (54). The functional regions of the global similarity network significantly enriched for interactions with VTI1 are indicated with blue dotted lines. Array genes enriched for novel differential negative interactions are shown in blue. The location of the VTI1 query gene on the global similarity network is indicated with a white node.
environmental perturbation, and thus conceptually resemble trigenic interactions, which involve three independent genetic perturbations, especially if a condition involves a drug with a highly specific cellular target. As a result, differential and trigenic interaction networks share several properties in common: (i) As observed for differential interactions, the average trigenic interaction degree for a given gene was correlated to its connectivity in the global genetic network (6); (ii) like novel differential interactions, trigenic interactions occur at similar reduced frequencies relative to digenic interactions in a reference condition ( fig. S7D); (iii) like modified and masked differentials, a substantial proportion of trigenic interactions overlapped with and exacerbated digenic interactions previously observed in the global digenic network (6); and (iv) as observed for novel differential interactions, novel trigenic interactions are also relatively rare and weaker but are highly coherent, in that they often involve array genes from the same biological process-however, they are also substantially more functionally diverse than digenic interactions (6).
GxE and GxGxE interactions involving the natural variation of different yeast strains can also be interpreted in the context of the global genetic interaction network and the properties of novel differential interactions. Using segregating populations of natural yeast isolates, recent surveys of GxG and GxGxE interactions between a deletion mutant and natural genetic variants showed that GxG interactions tend to connect genes within the same cellular function (42), whereas GxGxE interactions tend to be specific to a particular condition (43), and they often connected distantly related genes (44).
Quantitative trait loci (QTL) studies, focused on the fitness variation of different segregants derived from yeast crosses in different environments, showed that the majority of the phenotypic variation can be explained by singlelocus effects (GxE), whereas epistatic genetic interactions, analogous to GxGxE interactions, are relatively rare (4,45) and tend to affect only individuals with extreme phenotypes (5). On the basis of our analyses, most GxGxE interactions involve genes with single-mutant differential effects (GxE) (fig. S8E), which should be analogous to single-locus effects identified in QTL analyses and contribute to both the additive and epistatic variance; that is, the low epistatic variance often seen in QTL analyses may be partly explained by the partial contribution of interacting loci to the additive variance component (46). As a result, the remaining epistatic effect may only be visible in individuals with the most extreme phenotypes (5). Given that most GxGxE interactions overlap a genetic interaction in the global genetic interaction network, applying principles and gene modules inferred from a global network should allow for the detection of coherent epistatic interactions with relatively small effects that cannot be recovered by using QTL analyses (47,48).
We conclude that the yeast genetic interaction map derived from a single reference condition is highly robust to environmental influences because most connections on the global genetic interaction network remain unchanged or unmodified in a new environment. Although each new condition has the potential to map a relatively small number of novel differential interactions, the global digenic network mapped in a single condition is Costanzo 6. Functional distribution of novel differential negative interactions across conditions. (A) Novel differential negative interactions from each of the 14 conditions (x axis) were tested for enrichment for array genes grouped according to the biological processes indicated (y axis). Node size reflects the statistical significance of enrichment, and the shaded boxes along the diagonal indicate the biological process targeted by a particular condition. The average fold enrichment of novel negative interactions within biological processes targeted by a specific condition and interactions enriched within biological processes unrelated to the condition are shown in the box plot. (B) Regions of the global similarity network significantly enriched for array genes exhibiting novel negative differential interactions in (i) MMS or (ii) Monensin were mapped by using SAFE (54). The functional region of the global similarity network targeted by Monensin and MMS are indicated with blue dotted lines. Array genes enriched for novel differential negative interactions are shown in blue. Query gene(s) responsible for enrichment of novel differential interactions with the indicated array genes are shown as white nodes.
informative of differential interactions ( fig. S8, A and B) and should serve as an accurate and representative reference network. Nonetheless, differential interactions can reveal new functional connections between distantly related genes and bioprocesses. Given the significant association between the differential single-mutant fitness defect and the frequency of differential interactions for a given gene, a logical and efficient strategy may involve mapping differential interactions for specific genes required for normal growth in an environment of interest (such as stress condition, drug treatment, or microbial exposure).
Although environment has a modest impact, cell type-specific gene regulation has the potential to substantially complicate genetic network analyses in more complex systems. However, genome-wide studies suggest that only~8500 genes (~45 to 49%) are expressed in any given cancer cell line and that a "core set" of~7700 genes are expressed in the majority of all cancer cell lines examined to date (49). Thus, a systematic survey of genetic interactions based on a single cell line grown in one environment should be sufficient to map a global reference genetic network that encompasses this core set of expressed genes and provide a basic scaffold for the genetic wiring of a human cell. As observed in yeast, a global genetic interaction network for a model human cell line should reveal a hierarchy of functional connections among genes (1,11), and expansion of this network to specific cell types would enable the deciphering of mechanisms that underlie cancer cell-specific genetic vulnerabilities (50,51). Ultimately, reference genetic networks mapped in human cells should facilitate interpretation of allelic combinations of genes underlying inherited traits (3,47,48).

Methods summary Nonessential deletion and essential TS mutant arrays
The complete nonessential gene deletion and essential gene TS allele arrays were used for all single-mutant fitness and single-mutant differential fitness (GxE) analyses and are described elsewhere (11).

Diagnostic mutant array
Unless otherwise noted, genetic and differential interactions were based on analysis of query mutant strains crossed to a previously described, functionally representative diagnostic array comprising 1209 mutant strains (6). These included 1012 nonessential deletion mutant array (DMA) strains (labeled_dma#) and 197 essential TS allele array (TSA) mutant strains (labeled_tsa#). A complete list of diagnostic mutant strains used in this study is provided in data file 1 (20).

Query mutant strains
The set of 26 query mutant strains selected for this study were previously shown to have rich genetic interaction profiles that can recapitulate major bioprocess-enriched clusters on the genetic interaction profile similarity network and thus are representative of the global yeast genetic network (11,52). SGA query strain construction was conducted as described previously (53). A list of query mutant strains used in this study is provided in data file 1 (20).

Conditions
A list of 14 conditions and concentrations used in this study is provided in data file 1 (20).

SGA screening procedure
SGA experiments and selection steps were conducted as previously described (11,53), with the following modifications. To measure condition-dependent genetic interactions, every double-mutant array generated from a singlequery SGA screen was copied three times. One copy was grown in the standard SGA reference condition, whereas the two other copies were each grown in different conditional media ( fig.  S4). This configuration provided a matched reference control for every condition, which facilitated normalization of systematic experimental artifacts and improved accuracy of condition-specific genetic interactions measurements, as described below. Because every double-mutant array could be replicated a maximum of three times, each of the 26 query mutant strains was independently screened against the diagnostic array seven times, allowing us to measure digenic interactions in 14 different test conditions (2 test conditions + 1 reference condition/screen × 7 screens/query strain) (fig. S4). The entire screening pipeline was repeated twice, resulting in independent biological replicates for each query screen in every test condition along with 14 biological replicates of each query screen in the standard, reference SGA condition. We also repeated screens for a subset of four query mutant strains (data file 3) (20), in which double-mutant arrays were copied two times and each copy was grown in the standard SGA reference condition to measure the reproducibility of genetic interactions derived from matched reference control screens ( fig. S5C).

Single-mutant fitness scores
To derive accurate estimates of single-mutant fitness, we applied our colony size scoring method (19) to a set of control SGA screens, in which a query strain carrying a natMX marker inserted at a neutral genomic locus Costanzo   Many genetic interactions are modulated by the environment (white edges), creating a modified differential genetic interaction that varies in magnitude when compared with the equivalent genetic interaction on the reference network. Modified GxGxE interactions often connect functionally related query-array gene pairs. (D) GxGxE interactions (novel). Novel gene interactions, which were not detected in the reference condition, often connect genes involved in a bioprocess perturbed by a specific condition to distant, functionally unrelated genes.
was crossed to the kanMX-marked DMA (_dma#) and TSA (_tsa#) strain collections. Colony size measurements of SGA deletion and TS array mutant strains were based on an average of three replicate control screens conducted per each of 14 test conditions as well as the reference condition at 26°C. Colony size measurements were used to estimate singlemutant fitness in each condition as described previously (19), with the exception that bootstrapped means, instead of medians, across replicates were used in variance estimation and final fitness values. Because of technical reasons, single-mutant fitness for a small subset of mutant strains (0.15 to 1.0% of all strains) are not reported.

Single-mutant differential fitness scores
To obtain condition-specific fitness estimates, we computed the difference in colony size measured in a particular test condition versus the matched reference condition for each mutant. Single-mutant fitness and differential fitness estimates are provided in data file 1 (20). Because of technical reasons, single-mutant fitness for a small subset of mutant strains (0.15% to 1.0% of all strains) are not reported.

Genetic interaction score
To derive quantitative genetic interactions, we modeled colony size as a multiplicative combination of double-mutant fitness, time, and experimental factors as previously described (11,19). Briefly, for a double mutant carrying mutations of genes i and j, colony size C ij can be expressed as C ij = f ij · t · s ij · e, where f ij is the double mutant fitness, t is the incubation time, s ij is the combination of all systematic factors, and e is log-normally distributed random noise. The double-mutant fitness f ij can be further expressed as f ij = f i f j + D ij , where f i and f j represent the fitness of the two single mutants and D ij is a quantitative measure of the genetic interaction (genetic interaction score) between them.
Genetic interaction data corresponding to all tested gene pairs identified in the reference and 14 different conditions are provided in data file 3 (20). The data should be filtered before use. We suggest two different thresholds [intermediate (P < 0.05 and D j j > 0:08) and stringent confidence (P < 0.05 and D j j > 0:12)] that strike different balances between false negatives and false positives, as described in our previous studies (11,12). Although the majority of query mutant strains were screened in all test conditions, a smaller subset of query mutants could not be screened in all 14 conditions because of technical and data quality control reasons. In total, 22 of 26 of query mutants were screened in at least 12 test conditions, whereas five query mutants were screened for genetic interactions in less than 12 test conditions.

Differential interaction score
To score differential genetic interactions, genetic interaction scores derived from each condition were matched with a paired reference condition. Additionally, we used genetic interaction scores from the previously published reference genetic interaction network (11), termed "global" in the section below to normalize screen data before differential interaction scoring. The first step in scoring differential interactions was to apply a correction to each query, condition, and replicate screen that normalized the genetic interaction scores so that the corrected standard deviation for overlapping gene pairs matched the published genetic interaction network. We performed this per condition, per query, per replicate correction on reference (untreated) and condition (treated) scores as follows: correction c;q;r ¼ σ D untreated;c;q;r À Á σ D global;c;q À Á ε untreated;c;q;r ¼ correction c;q;r Â D untreated;c;q;r where c is the condition, q is the query, r is the replicate, ε is the uncorrected epsilon score, σ is the calculated standard deviation of the reference condition (untreated) epsilon scores for that query, condition (treated) and replicate, or the global epsilon scores for that condition and query from our previous work (11), andD is the corrected epsilon score. This conforms the variance of interaction measurements in each condition to a fixed reference and facilitates comparisons of interaction density across different conditions. Each condition (treated) network has a paired reference (untreated) network, which is used to compute the correction factor above. Three replicate scores were collected for every differential genetic interaction measurement. Some screens resulted in missing data for certain pairs because of data quality issues or strong fitness defects in the screened conditions. To mitigate the effect of missing data, we only considered interactions for which we had at least two replicate measurements. For interactions that had three replicates, we selected the two replicates with the highest correlation of corrected differential interaction scores for each query-condition pair.
We found that screening genetic interactions in compound stress conditions increased the variance in genetic interaction estimates. To mitigate this effect, we applied a variance stabilization correction to further correct conditional epsilon scores: where c is the condition, q is the query, r is the replicate,D is the corrected epsilon score, σ is the calculated standard deviation, andD is the variance-stabilized, corrected treated epsilon score. After applying these normalizations, the differential score is calculated for each of the paired replicates separately, as follows: differential score r ¼D treated;r −D untreated;r A final differential score, final untreated score, and final treated score are calculated from the mean of the two replicate differential scores, corrected untreated epsilon scores, and variancestabilized corrected treated epsilon score, respectively. Throughout these calculations, the standard deviations of the original measurements are propagated to derive an estimate of error on the final differential score. The resulting standard deviation is used to derive a P value for each differential interaction score, which was calculated as the two-sided probability of observing a more extreme score than the one measured, given a background normal distribution centered on zero with a standard deviation equal to the one observed. Differential interaction data corresponding to all tested gene pairs are provided in data file 3 (20). For analysis of individual interactions, we recommend that the data first be filtered before further analysis by applying one of our recommended thresholds (such as an intermediate threshold, P < 0.05 and D j j > 0:08).

Classifying differential interactions
Differential interactions were categorized as either novel, modified, or masked by comparing the genetic interaction scores for a given double mutant measured in a particular condition and the matched reference control (Fig.  2). Two additional categories of differential interactions were excluded from our analysis. First, gene pairs that did not show a significant genetic interaction (P < 0.05 and D j j > 0:08) in either a given test condition or the matched reference condition were excluded from further analysis because we could not be confident of a significant interaction in either the reference or the test conditions, individually, even though the differential score was significant. Second, double mutants that exhibited a significant and extreme negative genetic interaction (P < 0.05 and D j j < −0:3) in both a condition and the matched reference control were considered synthetic lethal in both the reference and test conditions and also excluded from further analyses of differential interactions.

Evaluating functional relations captured by differential interactions
We explored the functional information associated with different types of genetic interactions based on spatial analysis of functional enrichment (SAFE) neighborhood enrichment analysis (11,54). Query and mini-array genes were assigned to one of the 17 bioprocesses or neighborhoods on the global genetic interaction profile similarity map on the basis of SAFE analysis (11,54). For each query in each condition, we counted the number of interacting array genes that occurred in each of the 17 SAFE neighborhoods. For query-centric analyses (Fig. 5), we summed up interactions associated with a specific query gene across all 14 conditions and calculated the fold enrichment for each neighborhood with a one-sided Fisher's exact test, by comparing the number of interacting array genes assigned to a given neighborhood versus the total number of interactions with that particular query, using the fraction of array genes assigned to the same neighborhood across the mini-array as background. For condition-centric analyses (Fig.  6), we performed the same test by combining all queries within each tested condition. This analysis was performed for negative and positive genetic interactions (Figs. 5 and 6 and fig.  S9B), novel negative differential interactions (Figs. 5 and 6 and fig. S9A) using the miniarray screens, and novel negative differential interactions using the genome-wide screen on Benomyl condition ( fig. S9A).
A more detailed description of the experimental and computational analyses are provided as supplementary material.