Growth is both a system-level readout of cellular function and a quantitative trait. There are many considerations in quantifying growth, which is a complex function of cell proliferation, cell size, and viability/life span: each is potentially affected by genetic background (that is, gene deletions) and environmental factors (for example, temperature, nutrient conditions, drugs).

The phenotypic array method is based on analysis of scanned images of 8 × 12 cellular arrays spotted onto agar growth medium (Figure 2). In pilot studies, the image density of spotted cultures was linearly related to the number of cells (data not shown); however, imaging sensitivity was insufficient to measure exponential growth rates directly (see part A of the figure in Additional data file 1). Thus, we used the area under the growth curve (AUGC) as a unit measure of growth, representing the overall fitness (rate and total yield) for biomass accumulation (Figure 2).

The growth index (GI) predicts whether a deletion strain maintains growth proportional to the reference strain under a single perturbation. The equation is given in Figure 2i. It is formulated as a z-statistic, providing a standardized comparison of interactions. The rationale for each of the terms in the GI equation is as follows. First, the growth (AUGC) of each deletion strain under a perturbation is normalized to 'intrinsic growth' (AUGC of the same strain without perturbation) to account for effects of the deletion alone (see Figure 1, and Additional data file 2). Second, the mean and standard deviation of normalized growth for replicate cultures of the reference strain represent the effect of the perturbation alone and the experimental noise, respectively. Thus, a non-zero numerator signals a differential phenotypic response to the perturbation between the deletion strain and the reference strain, and the denominator normalizes to the experimental noise (Figure 2i).

Slow growth was found to be noisy; harsher perturbations therefore result in higher minimum GI values (Figure 2j). Also, strains with high GI values (antagonistic interactions) usually consist of intrinsically slow-growing strains having relative resistance but not better absolute growth than the reference strain (see Figure 1d, and Additional data file 2) 28. The GI may be modified to quantify the effect of a single perturbation (that is, gene deletion or drug exposure), whereby the AUGC, with and without perturbation, is substituted for the normalized ratios (for example, see column J in Additional data file 7). The GI, as a z-statistic, is normally distributed for replicates of the reference strain, with a mean value of zero (Figure 2j, and parts A and B in Additional data file 4), and is robust with respect to times after which growth has largely subsided (linear regression of GI values from the genome-wide HU screen calculated at 94 versus 123 hours had R² = 0.986, data not shown). Reproducibility of GI measurements obtained from independent experiments is represented in Additional data file 3.

A common method for recording growth phenotypes involves comparing serial dilutions of cells spotted onto agar 2930. Because we use a single dilution across a range of perturbations, and the AUGC is a fairly novel unit measure for growth, we examined the behavior of the GI with respect to the perturbation strategy (cell dilution versus drug dilution) (see Additional data file 4). Drug dilution and cell dilution were well correlated with respect to GI, but only drug dilution quantitatively delineates strong phenotypic interactions: for example, when drug dilution but not cell dilution permits growth (Figure 3).

The GI was used to screen for phenotypic interactions between gene deletion and three concentrations of HU (0, 50 and 150 mM). GI scores, for replicates of the reference strain, were normally distributed around a mean of zero with a range of -2.41 to +2.15 (Figure 2j). The distribution for deletion strains is typically bimodal: one mode with a mean of zero, like the reference strain, and one with a minimum value corresponding to zero growth (see Figure 2j, and Additional data files 4, 7). 'Sensitivity' to HU has been reported for yeast deletion strains identified from screens of sensitivity to gamma irradiation or the alkylating agent, methyl methanesulfonate (MMS) 3132. We found 'synergism' among many (75/118) of these previously reported genes, as well as many others (see Additional data file 8). Different methods of scoring growth and gene interaction, different growth media, and different HU concentrations may account for inconsistencies. There was only one deletion strain, kre22, which was found in both of the above studies but not in our screen. This proved to be a false-negative result in our screen when we sequenced the deletion barcode from our presumed kre22 deletion culture (see methods) 1733 and found a contaminating strain. We subsequently recloned the kre22 deletion and confirmed the synergistic interaction with HU in separate tests (data not shown).

Figure 2k depicts the relationship between effects of gene deletion alone and the interacting effects of gene deletion and treatment with 150 mM HU. On the basis of synergistic interaction, 274 strains were selected for further characterization, and 22 were selected on the basis of antagonistic interaction (Figure 2k, and see Additional data files 2, 8). Using the GI threshold of ±5.8, 94 strains met the GI criteria for both the 50 and 150 mM perturbations and 178 only for the harsher 150 mM HU condition. Only two interactions were scored as synergistic in the 50 mM HU screen alone. From Figure 2k note that antagonistic interactions (GI > 5.8) always occurred in the context of low intrinsic growth (see Figures 1d, 3d). This effect of 'alleviating' the deleterious effects of mutation was previously reported in transposon-mutagenized bacteria 28. Strains with extremely low intrinsic growth (AUGC < 600 on C media) were excluded from GI analysis (see Additional data file 7) the small denominator means relatively slight absolute differences in perturbed growth have a disproportionately large effect on the GI (see Figure 2i) in this setting, and are thus less reproducible. Among these excluded strains, there were at least two strains (gnd1 and ism1), which appeared growth-enhanced by the presence of HU; however, these were not studied further (see Additional data file 7).

In addition, the rnr3 and sml1 deletion strains were added to the HU-selected strains as negative controls in further studies, because while these two genes are known to regulate RNR activity 343536, they have small or undetectable effects on HU sensitivity.

The purpose of the GI is to screen for interactions using single perturbations (Figures 1, 2i). To validate its utility and further quantify the strength of the interactions, the 298 strains (with GI less than -5.8 or greater than 5.8) selected from the HU screen were perturbed with a range of HU concentrations inducing a wide spectrum of growth inhibition in the reference strain (Figure 3, and Additional data file 1). 'Phenotypic slopes' were calculated by regression analysis of AUGC, measured for each strain as a function of HU concentration (see Figures 1, 3, Materials and methods, and Additional data file 12). An 'interaction index' was formulated, essentially by substituting phenotypic slope values into the GI equation, yielding a z-statistic (see Figure 2i, Materials and methods, and Additional data file 12) where more negative, parallel, or less negative phenotypic slopes yield interaction index scores reflecting the probability of synergistic, additive or antagonistic interactions respectively (Figures 1, 3, and Additional data file 12). The reference range for the interaction index (replicate cultures of the reference strain) was -2.64 to +2.04, (see Additional data file 12). Using a z-score cut-off of ±5.8, 215 interactions were non-additive (197 were synergistic; 18 antagonistic), and 79 (27%) were additive (see Additional data file 12).

Genes with related functions had quantitatively similar interaction index scores. For example among the RAD52 epistasis group (MRE11-RAD50-XRS2, RAD51, 52, 54, 55, 57, 59, and RDH54) 37, all members showed equivalently strong synergistic interaction except for RAD59 (Figure 3a, and Additional data file 12), and RDH54, which functions in meiosis and was not selected from the HU screen 38. To confirm the weaker interaction of rad59, the respective homozygous diploid deletion strain was sporulated and dissected, the chromosomal deletion was confirmed by tag-sequencing (see Materials and methods), and multiple haploid segregants (of both mating types) were re-tested along with the homozygous diploid, which confirmed the phenotypic difference (data not shown). Thus, RAD59 is unique among homologous recombination genes in being relatively dispensable in the context of HU stress 39.

The interaction index for each of the 11 vacuolar H⁺-ATPase deletion strains was also uniform (see Additional data file 12), but in contrast to homologous recombination, the phenotypic effect of losing the vacuolar H⁺-ATPase was additive with the effect of HU (Figure 3b). The functional modularity seems to be due to the structural requirement of each subunit for complex assembly 40.

Several deletion strains required for vesicular protein trafficking displayed synergistic phenotypic effects (Figure 3c, and Additional data file 12). Modularity within this set of genes was suggested by the vacuolar protein sorting (VPS) genes, with near-exclusive recovery of class C and D VPS mutants, which displayed non-additive interactions, while the vps28 (the only 'non-class C/D' mutant) interaction was additive (see Additional data file 12). Furthermore, chc1, clc1 and end3, involved in plasma membrane to vacuole trafficking 40, shared similar interaction index scores (see Additional data file 12), further implicating vacuolar trafficking as an important module for buffering HU stress.

Eighteen strains exhibited antagonistic interactions (Figures 1d, 3d). Phenotypic slopes were relatively nonlinear in the context of antagonistic interactions, as absolute growth of deletion strains almost never exceeded reference-strain growth for a given perturbation (Figure 3d, and Additional data file 12: note R² values for strains with interaction index > 3).

There were nine strains (out of 4,852) with growth index greater than 5.8 or less than -5.8 at the 50 mM concentration only (see Additional data file 8, bolded values). The interaction index (see Additional data file 12) indicated these were spurious screening results due to experimental noise, either that associated with low intrinsic growth of the deletion strain and/or that of random experimental variation.

The analysis of phenotypic slope and interaction index given above verifies whether interactions are non-additive. It does not, however, insure that the interactions are biologically explained by deletion of the targeted gene. Unintended secondary mutations may occur during knockout transformation, and/or spontaneous growth-enhancing mutations may become fixed during strain propagation. To assess these possibilities, the genome screens were replicated in the homozygous diploid deletion set 41.

The comparison is summarized in Table 1 (see also Additional data files 7, 13, 14). Generally, the effect of homozygous deletion was more detrimental than haploid deletion for both intrinsic growth and HU resistance, consistent with previous data showing haploid deletion strains less affected by ionizing radiation and other DNA-damaging perturbations 31. From the 150 mM HU screens, there were 182 strains from both sets indicating synergistic interaction, 52 from the haploid but not the diploid set, and 213 from the diploid set only. Some deletion strain pairs could not be directly compared using the GI because either the haploid or the diploid deletion strain had inadequate intrinsic growth (less than 14% that of the reference strain) or the deletion was not represented in both sets.

In summary, the agreement between haploid and diploid screens supports the utility of the yeast deletion set and the GI for quantifying gene interactions. It appears that the haploid deletion strains are generally more fit than their homozygous diploid counterparts, whether perturbed by HU or not. It is possible that rapid accumulation of suppressor/adaptive mutations occurs on some haploid backgrounds, but given the small population size of each culture and the small number of selective generations, accumulation of suppressors seems an improbable general explanation for the increased fitness of haploid strains 4243.

The 298 deletion strains chosen from the HU screen were perturbed with four other drugs to determine the selectivity of gene interaction. Inhibitors of biologically diverse cellular functions were used: miconazole is a specific inhibitor of Erg11p, essential for ergosterol synthesis; cycloheximide is a specific inhibitor of Rpl28p, part of the large ribosomal subunit 44; cisplatin causes DNA intra- and inter-strand cross-links, resulting in DNA breaks; t-butyl hydroperoxide (TBHP) induces oxidative stress. Three concentrations of each agent were identified that inhibited reference strain growth equivalently to 50, 100 and 150 mM HU (see Additional data file 1), and all conditions were tested in parallel.

Selectivity for interaction with HU was high (see Additional data file 9). The correlation between GI value in response to HU and either miconazole, cycloheximide or TBHP was near zero (R² < 0.04). In contrast, correlation between cisplatin and HU was higher (R² = 0.13) (Figure 4, and Additional data file 5). The greater correlation between HU and cisplatin interactions presumably reflects overlapping pathways required for growth during perturbations to DNA replication 293132.

The experimental design and data structure of the phenotypic array are analogous to global gene-expression studies. The deletion strains are analogous to DNA hybridization probes for associating measurements with genes and the GI values represent directional (negative if synergistic, or positive if antagonistic) gene interaction quantities analogous to log₂ of mRNA hybridization ratios. Thus clustering algorithms can be used for global correlation of the direction, strength and selectivity of interaction 45. We refer to 'bio-modules' as sets of strains within a clustered phenotypic profile, indicating modular effects of the respective genes in responding to various growth perturbations (Figures 3, 4, 5, 6, 7).

GI data for the 298 HU-selected deletion strains, under all perturbation conditions, were clustered together (Figure 4, and Additional data file 10). The perturbations clustered by drug class rather than by inhibitory effect, indicating that the interacting effects of gene deletion are globally correlated by selectivity for the cellular target of each drug. If growth effects of drug inhibition and gene deletion were mostly additive, perturbations would cluster according to the degree of growth inhibition, as drug concentrations of equivalent effect were used (see Additional data file 1). HU and cisplatin perturbations clustered together with respect to the other perturbations, but separately with respect to each other (Figures 4, 5, 6, and Additional data file 6), reflecting the related but distinguishable biology of these two perturbations. As summarized below, we used gene annotations 464748 as a guide to understanding the biology of respective phenotypic modules.

The strains in cluster 2 (Figure 4) indicate strong synergistic effects, selective for HU and cisplatin perturbations. This cluster is enriched for DNA repair functions (Figure 5a), including recombination pathways 4950. The biological basis of this module probably relates to increased requirement for repair of single-strand DNA gaps and/or DNA double-strand breaks resulting from stalled replication forks as a consequence of either dNTP pool limitation or DNA cross-links 1920235152. Mre11p, Rad50p, and Xrs2p, along with the products of other genes in the RAD52 epistasis group, comprise a protein complex that repairs DNA breaks by homologous recombination. Rad51p has recombinase activity that is stimulated by Rad55p, and it is targeted to single-stranded DNA by Rad52p 3748. SGS1 (Bloom's Syndrome/RecQ helicase homolog), TOP3 (SGS1-interacting topoisomerase), MUS81, MMS4, and HPR5 have been characterized for their roles in detoxifying single-strand DNA lesions (see Figure 6b) 52. Within cluster 9a, five of six genes involved in sensing DNA damage (RAD24, RAD17, RAD9, DDC1 and MRC1) form a module, while MEC3 appears distinct, not interacting with cisplatin, and interacting antagonistically with other perturbations (Figure 4, cluster 9c, and Figure 6b). Only a few nucleotide-metabolism genes were identified from the HU screen. The interaction phenotypes of the strains ado1 (adenosine kinase; converts adenosine to AMP) and adk1 (adenylate kinase; converts AMP + ATP to ADP) were antagonistic, suggesting that the slow-growth phenotype of these strains could be due to basal substrate limitation. Thus, the growth-inhibitory effect of HU, observed at lower concentrations in the reference strain, would be masked until RNR activity becomes limiting (see Figure 1d) for dATP production. The strains apt1 (encoding adenine phosphoribosyltransferase, which converts adenine and phosphoribosyl pyrophosphate (PRPP) to AMP) and amd1 (AMP deaminase) indicated synergistic effects of gene deletion and HU perturbation, suggesting that these genes may have a regulatory function to compensate for growth deficiency, possibly by effecting an increase in the ADP substrate on demand. The hypothesis that ADP is the limiting substrate of RNR activity is consistent with the logic of negative feedback regulation, given that dATP is an allosteric inhibitor of RNR 22.

Threonine synthesis was another interesting module. It was the only amino-acid metabolism pathway recovered from the HU screen (see Additional data file 11). Deletion of HOM3 and HOM2 (homoserine is upstream of the threonine/methionine branch point) resulted in very strong and selective synergism with HU, whereas the deletion effects in downstream components (HOM6, THR1, THR4) were less selective (see Additional data file 5). AAT2 (aspartate aminotransferase), which is involved in the metabolism of aspartate, the substrate of Hom3p, showed weaker synergism (see Additional data file 12).

Genes required for normal mitochondrial function (SSQ1, TOM37, ATP5, RML2) also displayed HU synergism (see Additional data file 11). Reduced RNR function causes increased respiratory deficiency, possibly owing to perturbed mitochondrial DNA replication 53. Possibly connecting the threonine synthesis and 'mitochondrial' modules is the 'retrograde signaling' module (see Figure 7 and Discussion). RTG1, RTG2, RTG3 and MKS1 have been studied for their function in regulating the expression of genes of the TCA cycle 54555657 in response to mitochondrial dysfunction. Retrograde-deficient mutants are auxotrophic for aspartate, the precursor of threonine synthesis 54. MKS1, which interacts with RTG2 in retrograde signaling 56, had a similar interaction profile to the RTG genes (see part C of Additional data file 6).

The lst4 and lst7 (lethal with sec13) deletion strains, which clustered together, were phenotypically 'modular' with rtg2 as well (Figure 4, cluster 9a). LST4 and LST7 (along with an essential gene, LST8) have been shown to regulate transport of GAP1 (general amino acid permease) between the vacuole and plasma membrane 58. Furthermore, RTG2 signaling is negatively regulated by LST8 59. A hypothesis emerging from these interactions is that they represent a regulatory circuit whereby metabolic flux through threonine synthesis and degradation provides glycine, which may augment purine synthesis when RNR activity is compromised (see Discussion, and Figure 7) 4760.

A different bio-module was suggested by the broad phenotypic profile of interactions involving vesicular trafficking and vacuolar protein sorting genes 4061, and the strong synergism with cycloheximide (Figures 4 (cluster 3), 5b). A resulting hypothesis is that VPS genes are generally required for growth when protein synthesis is stressed (as with ribosomal poisoning by cycloheximide), and that protein synthesis is, in fact, generally stressed when cell growth is limited (see Discussion). Support comes from previous work showing morphological class C (no vacuole) and D (defective inheritance and acidification) vps mutants 62 to be most defective in proliferative responses of the endoplasmic reticulum, and sensitive to a diverse array of 22 perturbing conditions 30. Overall, only 12 of 46 vps mutants are in class C and D for vacuolar morphology 62; however, nine of 10 identified in this study were class C or D (the exception being vps28, which was phenotypically distinct in further tests, see Figure 3c). Our findings support a previous hypothesis that increased flux through the VPS pathway is a compensatory response required for maintaining growth in the face of a wide range of cellular perturbations 30.

The phenotypic profiles of the end3, chc1 (clathrin heavy chain) and clc1 (clathrin light chain) deletion strains were similar to the vps strains (Figure 7c, and Additional data file 12), implicating their shared role in trafficking to the vacuole, via endosomes 40. The strength of interaction for these mutants was similar to that of the vps mutants, further hinting at modularity in the requirement for flux through the vacuolar protein secretion pathway.

Tight phenotypic clustering was observed for strains carrying deletions of the structural subunits of the vacuolar H⁺-ATPase (VMA1, VMA2, VMA5, VMA7, VMA8, VMA6, VMA10, CUP5) (Figure 1 (cluster 5), see also Additional data file 10). Each of the subunits is essential for assembly and thus function of the complex 40. Overall, the interactions appeared weak and non-selective, suggesting that the result of losing the vacuolar H⁺-ATPase is additive (see Figure 3b).

An example of a protein complex that does not function quantitatively as a discrete biomodule is the GIM complex (PAC10, YKE2, PFD1, GIM5, GIM4, GIM3), which is involved in protein folding and maturation of tubulin and actin 6364. Only four of the six subunits (PAC10, YKE2, PFD1, GIM5) were identified in the 150 mM hydroxyurea screen (see Additional data file 11). Though similarities in the interaction profiles exist (Figure 6c), the relatively pleiotropic phenotypic effects of deleting genes of the GIM complex suggest that this complex is less modular than others in buffering growth against diverse perturbations.

The conclusions above about biomodules were based on enrichment of related genes within phenotypic clusters. However, genes of the same pathway may have opposite effects on pathway output, in which case phenotypic effects of gene deletion should not cluster together. For a more pathway-focused view, all genes were classified within broad categories of cellular function (gene expression, vesicular trafficking, cell polarization, DNA replication and cell metabolism) based on the literature (see Additional data file 11), and the subsets were individually clustered (Figure 6, and Additional data file 12).

The following broad conclusions drawn from the entire dataset (Figure 4) were confirmed. First, most strains deleted for known DNA replication genes interacted synergistically and selectively with the DNA synthesis inhibitors HU and cisplatin (Figure 6b,6c). Second, it appears that deletion of damage-sensing/checkpoint genes caused less synergism than loss of homologous recombination, particularly in response to cisplatin (Figure 6b). Third, genes involved in chromosome structure and movement showed quantitative effects comparable to the damage-sensing genes, but with less selectivity (Figure 6c). Fourth, strains defective in vesicular trafficking revealed interactions with many of the perturbations tested, the subset of vps deletions being notable for having more selective and stronger phenotypic interactions with HU and cisplatin (Figure 6a).

The pathway-oriented sub-analysis also highlighted clustering features that are less obvious when the entire dataset is analyzed (Figure 4). For example, Figure 6a recapitulates clusters 3 and 5 (from Figure 4), both composed of vesicular trafficking genes, while Figure 6b recapitulates clusters 2 and 9a, composed of genes required for DNA damage repair (Figure 6). It appears that these clusters have similar selectivity yet are quantitatively distinct as judged by strength of interaction. Overall, deletion of genes functioning in gene expression, cell polarization and cellular metabolism gave less modular phenotypic profiles, reflecting more pleiotropic interactions between deletions within these classes of genes as a function of changing cellular context (see Additional data file 12).

The MATa haploid deletion set was from Research Genetics (Huntsville, AL), release date 1/23/01. The reference strain, BY4741, is MATa, containing auxotrophic deletions of his3 leu2 met17 ura3 41. YM-1 liquid media was used for pre-growth of the yeast deletion set in 96-well plates and Hartwell synthetic complete agar medium was used for the growth arrays 78. HU, miconazole, cisplatin, TBHP, and cycloheximide came from Sigma. The Beckman Multimek 96, an automated 96-channel pipettor was used to inoculate, dilute and spot cultures. An Epson Expression 1640 XL scanner with A3 transparency unit was used to collect transmitted light images. For deletion strain confirmations, deletion cassettes were PCR-amplified, using the uptag1/downtag1 primer pair 41. PCR products were purified using the Qiagen PCR purification kit, and DNA sequencing was performed (dye terminator method) using KanB1 and KanC3 primers to identify the unique molecular barcodes 41.

All strains were inoculated (1:100) from thawed, glycerol stocks into new 96-well plates containing fresh YM-1 media + 2% glucose using the Multimek 96, and pre-grown at 30°C for 36-48 h. For spotting, pre-grown cultures were resuspended by orbital shaking before diluting 1:3025 (serial 55-fold dilutions) into water using the Multimek 96. Four microliters of the 1:3,025 diluted cultures were spotted (manually, using the Multimek) to synthetic complete agar medium containing 2% glucose, 2% agar, with or without perturbing agent (poured in omni-trays, Nunc). Agar plates were typically poured at least one day before use and 'dried' (incubated with the top off) for 30-60 min at 37°C before spotting cultures. Dilution of 1:3,025 yields approximately 250-500 cells per 4 μl with the reference strain, limiting the effect of growth measurement noise due to uneven initial distribution of cells, while increasing the number of generation times for detecting growth differences (data not shown). Multiple replicates of the reference strain were used to define experimental noise due to intra- and inter-experimental variations such as dilution, suspension, and spotting of cultures. The plates were scanned at designated times (see Image analysis).

Image analysis was performed on a Macintosh computer using the public-domain NIH Image program version 1.62, available on the Internet. Images were saved as 140 dpi black and white TIFF files. Up to 10 agar plates were included in a single scanned image. Scans were 'sliced' and 'stacked' to facilitate viewing and sorting data from each plate as needed for analysis. The images were then trimmed and modified to adjust for shadows, bubbles, and/or other artifacts containing pixels above the background intensity of agar. Each array (96 cultures) image contains a 600 × 400 pixel area, thus each culture in the 12 × 8 array is contained in a 50 × 50 pixel square. The average image density for each pixel of a culture, bounded by its square, is calculated using a 256-gray scale (0 = white), where values less than 90 are considered background and set to zero. The maximum pixel value within a 5-day reference strain culture spot was approximately 220; thus, maximum culture yield does not saturate the 256 grayscale. Growth curves were created for each strain by plotting image density vs time. The TrapZ function of MATLAB version 6 was used to calculate AUGC values for each growth curve, which were in turn used to calculate the GI. Microsoft Excel was used to manipulate and analyze data further. For hierarchical clustering of GI data, we used J-Express version 2.1 (with Euclidean distance measure and complete linkage) 79.

Phenotypic slope was determined by linear regression of AUGC vs HU concentration (mM) (see Figure 3 and Additional data file 12). For regression analysis, the first AUGC value less than 600 was the final value included; R² was greater than 0.95 (see Additional data file 12) for most strains. The interaction index was calculated like the GI (Figure 2i), substituting the corresponding phenotypic slope for each growth ratio (see Additional data file 12).

The gene classification was done subjectively, with reliance on the Saccharomyces Genome Database 47, Yeast Proteome Database 48, The Molecular and Cellular Biology of the Yeast Saccharomyces 46, and literature references found therein, for gene annotations.