Functional wiring of the yeast kinome revealed by global analysis of genetic network motifs

Synthetic dosage lethality screens of the kinome

We performed whole-genome SDL screens for 92 nonessential kinase deletion mutants (Supplemental Tables S1, S2), to examine the phenotypes of over 500,000 gene pairs (Supplemental Fig. S1; Sopko et al. 2006a). Dosage lethal interactions were quantified from colony size measurements, using the SGA score (Baryshnikova et al. 2010b). We applied a stringent score threshold (|ɛ| ≥ 0.2 and a P-value <0.05), which resulted in a high-confidence data set consisting of 934 dosage lethal interactions involving 69 different kinases (Fig. 1A; Supplemental Fig. S2A; Supplemental Table S3). To estimate false positive and false negative rates associated with our kinase SDL screens, a random subset of kinase–gene pairs (3000 pairs) that fell both below and above this cutoff were confirmed using an independent assay (re-transformation and serial spot assays; see Methods; Supplemental Table S4). Spot assays were also quantified and assigned confidence scores according to the calculated standard deviation (Supplemental Table S4). Also, interactions were only considered where the quantified growth difference was observed in the kinase deletion strains by eye. In total, we estimated a false discovery rate of ∼46% and a false negative rate of ∼50% (Supplemental Table S4). This analysis confirmed that the quality of our data set was comparable to other large-scale genetic (Baryshnikova et al. 2010b; Costanzo et al. 2010) and physical interaction data sets (Breitkreutz et al. 2010) and uncovered 410 additional SDL pairs that were not identified in our high throughput assay (Supplemental Table S4). In total, our SDL network represents >1300 interactions, of which ∼700 are confirmed by serial spot dilutions as a secondary confirmation assay (See Methods; Supplemental Table S4). Our data set represents the first systematic and quantitative assessment of dosage-sensitive interactions, expanding the roster of dosage-sensitive GIs for yeast kinases by ∼10-fold (Stark et al. 2006; Breitkreutz et al. 2008; Sharifpoor et al. 2011).

Figure 1.

Properties of the kinome SDL interaction network. (A) Distribution of SDL (blue), negative (red), and positive (green) GIs across the kinome using the stringent cutoff. The number of interactions for SDL-positive queries is plotted. (*) Kinases that were not screened for GIs. (B) Conditional screening of hog1Δ and cmk1Δ mutants expands the SDL interaction network. SDL screening of a hog1Δ strain was performed on 0.2 M NaCl (osmotic stress) and on 200 mM CaCl₂ for a cmk1Δ strain. (C) Representative serial spot dilution assays illustrating the SDL phenotype caused by overexpression of RPB2 in a hog1Δ strain specifically in activating conditions. (D) Bar graph showing the prevalence of positive (green bar) and negative (red bar) GIs at the intermediate SGA cutoff among kinase–gene (all), kinase–substrate, kinase–kinase, and kinase–kinase pairs that share the same target according to the KID gold standard, as a fraction of gene pairs screened. Prevalence of SDL interactions for each subtype is shown in parallel (blue bar).

Open in new tabDownload PowerPointLink to figure

We recently developed a literature-curated database for kinases (Sharifpoor et al. 2011), termed the Kinase Information Database (KID), which we used to produce a gold standard of 517 kinase–substrate (K–S) pairs (Supplemental Table S5; http://www.moseslab.csb.utoronto.ca/KID/). Since SDL has been used to identify targets of Pho85 (Sopko et al. 2006a), we assessed the number of known K–S pairs in the SDL interaction network. We were able to use 248 of the 517 gold standard pairs in our analysis since (1) many well-characterized substrates in the gold standard are targets of essential kinases that could not be screened using our standard SDL screening pipeline (e.g., >90 targets for Cdc28) (Sharifpoor et al. 2011) and (2) some known substrates were highly toxic when overexpressed or were absent from the array, preventing detection of an SDL phenotype. Of the 248 known K–S pairs that were tested in our SDL screens, 18 shared an SDL interaction (Supplemental Fig. S3A), which represents a highly significant 15-fold enrichment (P < 10⁻⁵³) compared to other nonsubstrate interacting partners of these kinases in the KID database (KID score <0) where no pairs were SDL. We repeated the analysis using another stringent negative control to avoid any bias. In this test, 248 K–S pairs were randomly shuffled to generate a list of nonsubstrate interacting partners (while preserving the degree distribution of the actual K–S network) and we still observed an enrichment of known substrates among SDL pairs (2.6-fold, P < 10⁻⁴). For example, Cla4 phosphorylates septins (Cdc10, Cdc12, and Shs1) during the G1/S transition of the cell cycle (Dobbelaere et al. 2003; Schmidt et al. 2003; Versele and Thorner 2004) and we identified two of the five septins (CDC10 and SHS1) in our SDL screen with a cla4Δ query strain. Also, the SDL screen for DUN1, a DNA damage checkpoint regulator, identified an SDL interaction with the ribonucleotide reductase inhibitor, SML1, which is targeted for degradation by Dun1-mediated phosphorylation during DNA damage (Supplemental Fig. S2A; Zhao and Rothstein 2002). Our analysis indicates that SDL interactions can be indicative of direct K–S relationships, emphasizing the use of this resource in identifying future targets of kinases. However, only a small portion of the kinase SDL network can be explained by known direct substrate relationships (1.3%), emphasizing that an SDL phenotype may also reflect other functional kinase–protein relationships in vivo.

We next assessed the biological relevance of the SDL network through comparison with other complementary large-scale data sets. Genes in the kinome SDL data set were significantly enriched for in vivo phosphoproteins (1.5-fold, P < 10⁻¹³; compiled in the PhosphoGRID database; Stark et al. 2010), regulated phosphoproteins in vivo (Bodenmiller et al. 2010) (1.2-fold, P < 0.05), and in vitro phosphorylated proteins identified by protein chip experiments (Ptacek et al. 2005) (1.5-fold, P < 10⁻⁵) (Supplemental Fig. S3B,C). While the SDL data set did not show significant overlap with physical interactions identified by traditional affinity purification and mass spectrometry approaches (Gavin et al. 2002, 2006; Krogan et al. 2006), we observed an enrichment for protein–protein interaction (PPIs) identified using two-hybrid (16-fold, P < 10⁻²⁵) (Ito et al. 2000, 2001; Yu et al. 2008) and protein overexpression–mass spectrometry experiments (fivefold, P < 10⁻⁶) (Supplemental Fig. S3D; Breitkreutz et al. 2010), which have recently been shown to detect binary and/or transient interactions (Yu et al. 2008; Breitkreutz et al. 2010), consistent with the transient nature of kinase signaling pathways. Overall, our results show that the SDL data set is not only enriched for direct substrates and phosphorylated proteins, but also for physically interacting partners of kinases and phosphatases, suggesting that the SDL network is functionally informative.

We found that the number of SDL interactions varied significantly for different kinases and for some kinases no SDL interaction was detected. We sorted the 69 kinases exhibiting at least one SDL interaction into two main groups, the “hub” SDL profile subclass, which is composed of nine kinases with >25 interactions, and a larger set of 60 kinases with a limited SDL profile, exhibiting between one and 24 SDL interactions (Fig. 1A; Supplemental Fig. S2A). Hub kinases often coordinate cell cycle progression and cell polarity (Cla4, Elm1, Gin4, Hsl1, Pho85, Bud32, Kin4) or regulate cell wall integrity (Bck1 and Slt2) (11-fold, P < 10⁻⁴) and their mutants tend to have phenotypic defects that are detectable in standard growth conditions (e.g., elongated buds, flocculated colonies, aberrant cell wall involving morphologies and fitness defects) (Martin et al. 1996; Edgington et al. 1999; Asano et al. 2006; Breslow et al. 2008; Watanabe et al. 2009), reflecting their requirement for normal cell cycle progression. Consistent with these observations, the hub kinases were also significantly enriched for interactions with genes involved in cell cycle progression (11-fold, P < 10⁻⁴) (Supplemental Fig. S2B), which suggests that high connectivity in the SDL network may reflect a requirement for active kinases throughout the cell cycle (Bodenmiller et al. 2010).

Condition-specific SDL screens reveal expanded kinase SDL profiles

The SDL interactions derived from the hub kinase screens alone overlapped with gene pairs co-annotated to the same biological functions as frequently as the complete set of SDL interactions identified from screening 92 kinase mutants (Supplemental Fig. S2C). This observation suggests that almost all of the functional connections associated with the SDL interactions identified are derived from the hub kinase genes and that most kinases may only be activated under specific genetic or environmental conditions. To address this hypothesis, we performed SDL screens for a subset of kinases under gene-specific biological conditions.

We first screened a hog1Δ kinase mutant strain for SDL interactions in the presence of salt stress. Hog1 is required for the high osmolarity glycerol response and translocates to the nucleus to activate gene expression in response to osmotic stress (Reiser et al. 1999). HOG1 failed to show any SDL interactions under standard growth conditions (Fig. 1B). In contrast, a hog1Δ screen in the presence of 0.2 M NaCl identified 78 SDL interactions (Fig. 1B; Supplemental Table S6). The salt-specific hog1Δ SDL profile was enriched for genes involved in transcription and chromatin remodeling (P = 1.2 × 10⁻⁴), consistent with known features of Hog1, which include binding to the RNA polymerase subunit Rpb2 at gene promoters (Supplemental Fig. S4A; Alepuz et al. 2003; Mas et al. 2009). Spot dilution assays were used to confirm all SDL interactions, including Rpb2 (Fig. 1C).

A second condition-specific SDL screen with a calmodulin-dependent protein kinase, CMK1, also uncovered relatively few interactions in standard growth conditions. CMK1 has reported roles in stress response and is sensitive to calcium levels (Pausch et al. 1991). We identified 11 new SDL interactions when the CMK1 SDL screen was conducted in the presence of calcium (Fig. 1B; Supplemental Table S6) and the confirmed SDL interactions were enriched for factors including vesicle-mediated transport genes required for calcium signaling (P = 5 × 10⁻³) and calcium-responsive proteins (Supplemental Table S6; Supplemental Fig. S4B). Thus, our results suggest that the apparent resistance of many kinase mutants to gene overexpression reflects the requirement for kinases only in certain activating conditions and that conducting screens in a variety of biologically relevant conditions will substantially increase our view of kinase biology, particularly regarding kinase dosage sensitivities.

Loss-of-function SGA genetic interaction screens of the kinome

In addition to environmental cues, SDL interaction degree is likely dependent on the extent of functional redundancy shared between different kinases. Thus, we also assembled genome-wide surveys of the kinase genes that displayed SDL profiles for loss-of-function SGA GIs. We included GI data for 45 kinases from Costanzo et al. (2010) (37 of which provided SDL interactions) and performed genome-wide SGA screens for an additional 30 kinase queries (Supplemental Table S7; Supplemental Fig. S5). In total, we identified 2789 unique negative and 203 positive high-confidence interactions (SGA score: ɛ < −0.12, ɛ > 0.16, P < 0.05) (Supplemental Table S7). Many different kinase mutants displayed a relatively large GI profile, revealing the genetic conditions under which these kinases are required for cellular fitness. For example, while HOG1 failed to show any SDL interactions under standard growth conditions, it showed a substantial number of negative SGA GIs (Supplemental Figs. S2, S5).

In a previous study, Fiedler et al. (2009) assessed GIs among a selected subset of ∼400 genes encoding kinases, phosphatases, and potential pathway components. They report a bias toward the occurrence of positive GIs between kinases and their substrates. We observed a similar trend using our genome-wide approach and a new gold standard list of K–S pairs (Fig. 1D). Our results also indicate that both positive and negative GIs are enriched amongst kinase–kinase (K–K) pairs, including those that share the same target (Fig. 1D).

To address whether K–K redundancy was contributing to the lack of an SDL phenotype, we assessed the number of positive and negative GIs amongst K–K pairs, including those that share the same substrate. We found that the ratio of positive to negative GIs remains unchanged for K–K pairs (P = 0.45, Fig 1D) when being compared with kinase-any pairs. In fact, negative interactions between K–K pairs and between random gene pairs occurred at the same frequency when assessed across the genome (∼2%) (90 negative interactions among 4359 unique tested pairs). Also, of the 99 K–K pairs in the SGA data set that share at least one substrate in our K–S gold standard (Supplemental Table S5; Sharifpoor et al. 2011), 10 kinase pairs showed a negative (P < 10⁻¹²) GI with each other, while four pairs displayed a positive interaction (P < 10⁻⁴), corresponding to a seven- and sixfold enrichment, respectively. This analysis suggests that K–K pairs in general can have both positive and negative GIs (Fig. 1D), including those that share the same target. Consistent with a complex model of GIs between K–S and K–K pairs, we found that genes encoding proteins that physically associate with kinases may positively (threefold enrichment, P < 0.005) or negatively interact with the kinase (threefold enrichment, P < 10⁻⁴) (Breitkreutz et al. 2010). Together, our analysis shows that redundancy among kinases at a digenic level is not a major contributor to the buffering effect of kinase pathways and that K–K pairs regulating the same target can have either positive or negative GIs with each other. Interestingly, when we assessed the patterns of K–S and K–K enrichments for SDL interactions, we found that not only are K–S pairs more likely to show an SDL interaction than random, but also SDL interactions are highly enriched amongst K–K pairs including those that share the same substrate (Fig. 1D). Our combined analysis suggests a highly complex model of regulation in kinase pathways that requires more detailed network analysis for understanding kinase interactions.

Integration of the SDL and SL data sets produces a functionally informative meta-network

Consistent with other work (Kelley and Ideker 2005; Lee et al. 2008), we reasoned that integration of our kinase data sets should enable identification of regulatory motifs that provide insight into the complexity of kinase pathways and kinase function. This is particularly interesting in the context of SDL, since many SDL interactions could not be explained by literature-curated K–S relationships. Therefore, we first leveraged the SDL interaction network with the complementary SGA data set to produce a GI meta-network. Next, we super-imposed our literature-curated K–S data set onto our meta-network to search for triplet sub-networks involving all possible combinations of SDL and double deletion mutant interactions (See Methods). In total, the meta-network was based on 69 SDL and 75 SGA kinase screens and includes 1344 dosage interactions, 7427 double mutant interactions, and 517 literature-curated K–S pairs (Sharifpoor et al. 2011). We also included 71,886 interactions among gene pairs if the pair was one-hop away from a kinase in the global GI network (Costanzo et al. 2010). We applied an automated network motif detection algorithm to identify three-gene motifs that are significantly overrepresented in our integrated genetic network. Triplet motifs derived from our integrated network were statistically enriched compared with motifs derived from randomized networks (P < 0.05). We identified 2685 motifs consisting of 12 different combinations of SDL, SGA, and literature-curated interactions (Fig. 2; Supplemental Fig. S6; Supplemental Table S8). We reasoned that each motif in this resource may be associated with a specific biological interpretation from which mutant phenotypes and regulatory relationships can be inferred. Therefore, we tested three of the five most common motifs for biological relevance.

Figure 2.

Overrepresented SDL motifs identified through combined analysis of SDL, SL/SSup GI data sets, and gold standard kinase–substrate pairs. Five types of sub-networks were enriched against the randomized data set: (A) counteracting; (B) balancing co-regulation; (C) converging regulation; (D) diverging regulation; and (E) SDL cascades. The number and fold-enrichment of triplet gene pairs identified are listed, as well as examples of known and novel motifs for each group. Possible mechanistic models describing each motif are highlighted. Some motifs fall under more than one subtype.

Open in new tabDownload PowerPointLink to figure

First, we examined what we have dubbed the “counteracting” motif category which involves two SDL interactions connected by a known K–S pair, where the substrate may also be another kinase, such as an interacting downstream kinase within a cascade (Fig. 2A, 5.8-fold). For example, Bck1, a MAP kinase kinase kinase (MKKK), is connected indirectly to the cell wall integrity MAP kinase (MAPK) Slt2 through the Mkk1/2 MAP kinase kinases (MKK). Both Bck1 and Slt2 showed SDL interactions with an uncharacterized gene YMR074C, encoding a protein with homology with human PDCD5, which is expressed in tumor cells during apoptosis (Fig. 3A; Hong et al. 2009). A simple model for SDL interactions predicts that Slt2 may negatively regulate the function of Ymr074c, perhaps through direct phosphorylation. Alternatively, Ymr074c may act in opposition to inhibit the cell wall integrity pathway (Fig. 3B). In tests for a functional connection between Slt2 signaling and Ymr074c, we found that deletion of YMR074C partially suppressed the fitness defect of both bck1Δ and slt2Δ deletion alleles in the presence of caffeine, a stress-inducing reagent (Fig. 3C). Importantly, we tested all 30 genes that showed a similar pattern of interaction with bck1 and slt2 for sensitivity to caffeine and a second cell wall stressor, Calcofluor White (CFW), and found that deletions of six genes (SNF7, COQ1, KRE6, PEP7, CHO2, SBE22; Supplemental Table S9) suppressed the sensitivity of bck1Δ and/or slt2Δ strains to cell wall damaging agents (twofold enrichment, P < 0.03). Thus, the motif produced by analysis of our integrated genetic network was highly predictive of the mutant phenotypes of kinase regulated genes.

Figure 3.

YMR074C is an inhibitor of the Slt2-MAP kinase cell wall integrity pathway. (A) Overrepresented triplet motif that identified YMR074C to be SDL with two kinases of the cell wall integrity pathway, Bck1, and its downstream target Slt2. (B) Model describing the role of Ymr074c as an inhibitor of cell wall integrity. (C) Overnight cultures of bck1Δ, slt2Δ, ymr074cΔ, single or double mutants were serially diluted and spotted onto YPD medium in the absence (left) and presence (right) of 10 mM caffeine. (*) Phenotypic rescue.

Open in new tabDownload PowerPointLink to figure

We refer to another significantly enriched motif as “balancing co-regulation” (Fig. 2B, 16,000-fold, P < 10⁻¹⁰). This motif involves a gene connected to both a kinase and one of the kinase substrates by an SDL and a negative GI, respectively. We suspect that this motif may identify proteins or protein complexes regulated by the same kinase, but in different ways (positive versus negative regulation). For example, one “balancing co-regulation” motif involved the Cla4 kinase, which is required for phosphorylation and proper localization of Lte1, a surface-localized mitotic exit regulator (Hofken and Schiebel 2002; Seshan et al. 2002; Yoshida et al. 2003). CLA4 showed an SDL interaction with BEM1, which in turn had a negative GI with LTE1 (Supplemental Table S7). Bem1 also physically interacts with Cla4 (Gulli et al. 2000; Bose et al. 2001) and this complex is required for phosphorylating Cdc24 (the guanine nucleotide exchange factor for the master regulator of cell polarity Cdc42), a modification that ultimately inhibits bud growth. The CLA4-LTE1-BEM1 balancing co-regulation motif (Fig. 2B) suggests that Cla4 inhibits its target, the Bem1–Cdc24 complex, but activates Lte1 in the mitotic exit network, thereby coordinating cell polarized growth with cell division. In this case, our motif analysis accurately identified known regulatory relationships in a complex pathway controlling mitotic exit, validating the approach. We further predicted phenotypes for genes belonging to seven “balancing co-regulation” motifs involving the cell wall integrity kinases Bck1 and Slt2 and found that deletion of four out of seven tested genes (SMY1, EDE1, PEX19, CHO2; Supplemental Table S9) suppressed the sensitivity of bck1Δ and slt2Δ mutants to cell wall damaging agents (5.5-fold enrichment, P < 0.003), suggesting that these genes are likely involved in cell wall integrity (Supplemental Table S9).

We called a third highly enriched motif “converging regulation” (Fig. 2C), in which a kinase pair, or a kinase–gene pair, negatively interact and both share an SDL interaction with a third gene. This motif class may identify kinases or enzymes that act synergistically to control a biological outcome through the same gene. Alternatively, kinases involved in the same motif may regulate each other to then influence a downstream target. We explored converging regulation motifs by testing regulatory relationships predicted by another motif involving Cla4. As noted earlier, Cla4 is required for phosphorylation and proper localization of Lte1 in vivo during mitotic exit (Hofken and Schiebel 2002; Seshan et al. 2002; Yoshida et al. 2003). Localization of Lte1 to the cortex also requires its physical association with the scaffold protein, Kel1 (Seshan et al. 2002). We identified a converging regulation motif in which the genes encoding Cla4 and Hsl1, another cell polarity kinase, showed a negative GI with each other and shared an SDL interaction with KEL1 (Fig. 4A). This motif led us to make two predictions: (1) Cla4 and Hsl1 have a synergistic and direct role in regulating Kel1; (2) Hsl1 may have a previously unappreciated role in Kel1 regulation; since the primary role for Kel1 during the cell cycle is to anchor Lte1 to the cortex (Hofken and Schiebel 2002; Seshan et al. 2002), it follows that Hsl1 may regulate the association between Kel1 and Lte1. We tested these predictions by analyzing Kel1 protein and its association with Lte1 in cla4 and hsl1 mutant strains (Fig. 4B,C). First, we analyzed Kel1 protein by Western blot and saw reduced Kel1–TAP fusion protein in both hsl1Δ and cla4Δ deletion mutant strains (Fig. 4B), but not in a strain deleted for gin4Δ, another budneck kinase that is also SDL with KEL1 (Supplemental Fig. S7). Slow migrating isoforms of Kel1 were also reduced in the cla4 mutant, consistent with a kinase–substrate relationship (this could not be assessed in the hsl1 mutant, since protein levels were so low). Second, we used coimmunoprecipitation to show that the interaction between Kel1 and Lte1 (Hofken and Schiebel 2002; Seshan et al. 2002) was substantially reduced in an hsl1Δ mutant but not a gin4Δ mutant, indicating a requirement for Hsl1 in Kel1 function (Fig. 4C). Taken together, these data confirm the predictions of the converging regulation motif and suggest that Hsl1 and Cla4 may function together to regulate Kel1 function in mitotic exit.

Figure 4.

Kel1 protein is co-regulated by budneck kinases. (A) Overrepresented motif (gray triangle), supporting a functional connection between Kel1 and the two budneck kinases Cla4 and Hsl1 in a converging regulation model. (B) Endogenous Kel1 protein levels are reduced in hsl1Δ and cla4Δ but not in gin4Δ mutant strains. Kel1-TAP protein was immunoprecipitated from the indicated budneck kinase mutant strains and Western blots were probed with anti-PAP to detect total Kel1 protein. Electrophoretic mobility shift of Kel1 protein is shown (Kel1-P). Swi6 was used as a loading control. (C) Integrated Kel1-TAP protein was immunoprecipitated in the presence of Lte1-Flag and association of Lte1 with Kel1 was detected using α-flag antibody. The amounts of Lte1-Flag in the whole cell extract (WCE; top panel) and in the Kel1 immunoprecipitate (bottom panel) are shown. (Middle panel) The amount of Kel1-TAP in the immunoprecipitate (anti-PAP). Immunoprecipitation assays were performed in wild-type, hsl1Δ, gin4Δ background strains. cla4Δ strain was not tested, due to severe toxicity upon Kel1-TAP expression. (D) Model for regulation of Kel1 by budneck kinases. Lte1 is phosphorylated by Cla4, which triggers binding to its anchor, Kel1. Hsl1 regulates the stability of Kel1 protein, which allows the anchor to become available for binding to Lte1-P. Together, the complex inhibits mitotic exit possibly to delay budding to allow the cell to reach the correct size. Later, Cdc14 phosphatase releases Lte1 from its anchor in the cortex into the cytoplasm in order to activate mitotic exit in large budded cells.

Open in new tabDownload PowerPointLink to figure

Synthetic dosage lethal screens

The SDL screening protocol was adapted from Sopko et al. (2006a) and is summarized in Supplemental Figure S1 (Sopko et al. 2006a). Refer to the online Supplemental Materials and Methods for detailed SDL screening and confirmation protocol (http://andrewslab.ccbr.utoronto.ca/data/). In brief, a collection of yeast strains deleted for query kinases was constructed in a strain background compatible with SGA analysis (Supplemental Tables S1, S2; Costanzo et al. 2010). Kinase deletion mutants (marked with the NAT^R cassette) were crossed to a GST-ORF overexpression array (Supplemental Table S2; Sopko et al. 2006a; Costanzo et al. 2010) using the SGA protocol (Tong et al. 2001). Screens were performed in a 1536 colony format, where each colony is represented four times on the array, producing eight replicates per gene. Gene overexpression was induced by pinning the final haploid array onto medium containing 2% galactose and plates were then incubated at 30°C for 2 d. Colony sizes were quantified using an adapted SGA protocol (Baryshnikova et al. 2010a). SDL and SDS interactions were scored using the previously described SGA score method with the assumption that no kinase queries have significant fitness defects on the final selection plates. SDL and SDS interactions were defined as gene pairs that satisfy the following SGA score cutoff: |ɛ| ≥ 0.2, P < 0.05 (Supplemental Table S3). To estimate false discovery rate, we confirmed >3000 interactions by direct transformation both in wild-type and mutant backgrounds followed by serial spot assays where a difference in colony size was observed in the kinase deletion strains by eye. Spot assays were quantified and assigned confidence scores according to the calculated standard deviation (Supplemental Table S4). In total, we estimated a false discovery rate of ∼46% and a false negative rate of ∼50% (Supplemental Table S4), which is comparable to rates observed for other large-scale genetic (Baryshnikova et al. 2010b; Costanzo et al. 2010) and physical interaction data sets (Breitkreutz et al. 2010). Conditional SDL screens were performed as above but final haploid arrays were pinned onto the medium containing the appropriate inducer at a concentration reported to be consistent with viability. Kinase mutants with mating defects (STE7, STE11, STE20) or those otherwise refractory to the SGA protocol (BUB1, IRE1, SSN3, SSK22, SCH9) were excluded in this study. We were unable to make query kinase deletion strains for 10 kinases for technical reasons (Supplemental Table S1).

Negative and positive genetic interaction screens

For analysis of growth defects in double deletion mutants, the kinaseΔ::NAT^R (Supplemental Table S2) query strains were crossed to the viable haploid deletion array (Giaever et al. 2002; Baryshnikova et al. 2010a) as previously described (Baryshnikova et al. 2010a). Negative and positive SGA GIs were quantified as described elsewhere (Baryshnikova et al. 2010b). An intermediate SGA score cutoff (|ɛ| ≥ 0.08, P < 0.05) was applied to SGA data as recommended elsewhere (Baryshnikova et al. 2010b; Costanzo et al. 2010), unless otherwise indicated. We confirmed a subset of double mutant interactions between K–K pairs identified in our screens by tetrad dissection, and automated liquid growth curve assays.

Literature-curated kinase data set (KID)

We compiled a highly detailed literature-curated database of K–S interactions for all kinases (Kinase Information Database; KID) (Sharifpoor et al. 2011). Biological interactions involving kinase interactions were compiled by a group of eight kinase experts and tagged with their associated PMIDs. KID is a ranked database for kinase interactions that scores K–S pairs by applying likelihood ratios using a known gold-standard K–S list and comparing it to randomized kinase–gene pairs (Sharifpoor et al. 2011). The ranked score on KID is defined as a likelihood ratio of the prevalence of each experimental category in the gold standard K–S set compared with a randomized list of kinase–gene interactions (negative bin). We considered the gold standard K–S pairs as the 517 pairs reported on KID corresponding to the stringent cutoff (P ≤ 0.01) (Sharifpoor et al. 2011). The literature-curated gold standard list of K–S pairs is described in Supplemental Table S5.

Analysis of triplet motifs using double mutant and dosage interactions

We used Fanmod (Wernicke and Rasche 2006) to identify overrepresented three-edge sub-networks in the combined meta-network of (1) literature-curated K–S interactions, (2) the combined confirmed and unconfirmed SDL interactions above our defined cutoff, (3) negative GIs (bidirectional edges), and (4) positive GIs identified both as query genes from our study or as array hits from a previous analysis of fitness defects in double deletion mutants (Supplemental Table S6; Costanzo et al. 2010). Fanmod is a fast network motif detection tool to identify connected subgraphs that occur significantly more often than in random networks. Crucially, Fanmod allows for motif detection in colored networks (i.e., those with multiple edge types). We ran Fanmod with a maximum subgraph size of three (triplet motifs) and four edge types as described. In the first step, Fanmod enumerated all subgraphs of size three and grouped them into isomorphic subgraph classes (Wernicke and Rasche 2006). Next, Fanmod determined the frequency of subgraph classes in randomized graphs. Two-hundred random graphs were generated from the original network by switching edges between vertices (exchanging only edges of the same type) while preserving the number of bidirectional edges (we used this randomization scheme because our combined network is directed). Finally, significance of each subgraph in the network was calculated and those with P < 0.05 were further considered. Significant triplet motifs were extracted using a program implemented in MATLAB that performs an exhaustive search over all possible gene combinations for those fitting each motif. Starting with a list of pairs that fit one line in a triangular motif (e.g., A→B; K–S pair), the script searches through each given rule generated in Fanmod to identify a third gene that relates to the first pair according to any of the stated rules.

Statistical analyses

Overlap of synthetic dosage lethal interactions with other data sets was analyzed as follows. For each pairwise interaction data set (e.g., known kinase–substrate pairs, protein–protein, or GIs) we first assembled a list of gene pairs that have been screened for SDL interaction (either showing or not showing SDL) and for which interaction information from the other data set was also available (e.g., either being a kinase–substrate pair or not displaying kinase–substrate interaction). For kinase–substrate interactions, we defined noninteracting K–S pairs as those which have a negative likelihood ratio score in the KID literature-curated database. For protein–protein interactions, we assumed that nonreported protein pairs do not interact. Next, we computed the significance of the overlap between the SDL and the other interaction data set using χ² test of independence. Overlaps between loss-of-function genetic interactions (negative and positive GI) and kinase–kinase or kinase–substrate pairs were also investigated in a similar manner. To test the overlap of SDL hits (i.e., list of overexpressed genes that show at least one SDL interaction with a kinase deletion strain) with other gene lists (e.g., list of genes whose protein products are phosphorylated in vitro) we also performed χ² tests of independence.

Other yeast strains and plasmid construction

Kinase deletion strains carrying a TAP-tagged allele of KEL1 were constructed by crossing a MATa KEL1-TAP-HIS3 strain (Ghaemmaghami et al. 2003; Costanzo et al. 2010) to the MATα kinaseΔ::NAT^R query strains (Supplemental Table S2). Kinase–gene double deletion strains were constructed by crossing NAT^R kinase deletion mutants (see above and Supplemental Tables S1, S2) to the kinase deletion strains from the yeast deletion collection (Giaever et al. 2002) (KAN^R gene deletion strains) and dissecting tetrads. MATa haploids of the desired double mutant genotype were selected for analysis by replica plating segregants onto YEPD media (1% yeast extract, 2% bactopeptone, 0.004% tryptophan, 0.004% adenine, 2% glucose) containing cloNAT and G418.

A LEU2-based plasmid expressing a C-terminally Flag-tagged allele of LTE1 was constructed by cloning LTE1-HA from the mORF collection (Gelperin et al. 2005) into a pRS315 low copy CEN vector using the Invitrogen Gateway cloning technology (Supplemental Table S2). First, LTE1-HA from the mORF collection was transferred to pDONR-221 plasmid, which was then used to transfer the LTE1 cassette into the recipient plasmid pRS315-FLAG-LEU2. The resulting construct, pRS315-LTE1-FLAG-LEU2, was then transformed into the indicated kinase deletion strains in the presence or absence of integrated KEL1-TAP for coimmunoprecipitation assays (see below).

Serial spot dilutions

Standard methods and media were used for yeast growth and transformation. Spot assays were performed by growing yeast cultures in appropriate media to saturation and spotting equal volumes of 15-fold serial dilutions (5 μL) onto plates with appropriate supplements to assess growth. All glucose-containing plates were incubated for 2 d, while galactose-containing plates were grown for 3 d at 30°C. To assess cold sensitivity, YEPD plates were incubated for 7 d at 15°C. To test strain sensitivity to cell wall stressors, we included 10 mM caffeine or 64 μg Calcofluor White in YEPD plates and grew cells for 4 d.

Western blot and coimmunoprecipitation analysis

To prepare samples for Western blot analysis, 50 mL of cells were grown in the appropriate media to an optical density (O.D.₆₀₀) of 1.0 and washed once with phosphate buffered saline (PBS). Samples were resuspended in 250 μL of solution containing 1% β-mercaptoethanol in 0.25 N NaOH and incubated on ice for 10 min; 160 μL of 50% TCA were added and samples were incubated for another 10 min on ice. After a 10-min centrifugation at high speed, the precipitated protein was isolated and resuspended in 100 μL of 2× SDS sample buffer. Samples were resolved on 6% SDS polyacrylamide gels and transferred to polyvinylidene difluoride (PVDF) membranes following standard procedures for Western blot analysis. Peroxidase anti-peroxidase complex, α-PAP (P3039; Sigma Aldrich), was used to detect TAP-tagged proteins. Flag-tagged proteins were detected using α-Flag M2 (Sigma Aldrich) and α-GST primary conjugate (GE Healthcare) was used to assess migration changes for GST-tagged proteins. Swi6 protein levels were assessed as a loading control using an affinity purified α-Swi6 antibody (Sidorova and Breeden 1993). For analysis of coimmunoprecipitation of tagged Lte1 and Kel1, overnight cultures of BY4798 (MATα ura3Δ::NAT^R KEL1-TAP-HIS3), BY4800 (MATα hsl1Δ::NAT^R KEL1-TAP-HIS3), and BY4802 (MATα gin4Δ::NAT^R KEL1-TAP-HIS3), carrying either the pRS315-LTE1-FLAG or control plasmids, were grown overnight in 2% raffinose and expression of LTE1 was induced by addition of 2% galactose for 9 h. Cells were lysed in Lysis Buffer (50 mM Tris-HCl, pH 7.5, 100 mM NaCl, 1 mM EDTA, 5 mM NaF, and protease inhibitors) and samples were clarified by centrifugation at high speed for 10 min. Total protein was incubated with IgG sepharose (Amersham Biosciences) overnight at 4°C with gentle shaking. IgG resin was washed three times in lysis buffer and resuspended in 2× sample buffer for Western blot analysis.