An Important Percentage of Yeast Essential Genes Affect Fitness When Duplicated.

We measured the fitness effect of gene duplication using high-resolution competition assays with fluorescently labeled cells in coculture (Fig. 1A) (23). We individually duplicated 899 essential genes using single-copy centromeric plasmids (pCEN) expressing the genes under their native promoter and 3' untranslated region (24). In parallel, we generated a distribution of control strains in which we competed a wild-type (WT) strain (also used as a reference) with itself. Each of the 192 replicates of this control set is an independent colony from a transformation.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 1.

More than 10% of yeast essential genes affect fitness when duplicated. (A) Relative fitness was measured using a high-resolution competition assay (23). We cocultured a mCherry-tagged strain carrying an extra copy of an essential gene on a centromeric plasmid with a CFP-tagged reference strain carrying a control plasmid. We followed the ratio of the two populations for up to 28 generations to calculate a slope, which corresponds to the selection coefficient (s). (B) Cumulative distribution of selection coefficients of all the 899 strains tested (Dataset S1). Each dot represents a strain expressing an additional gene copy. The black dots represent the distribution of 192 biological replicates of reference-versus-reference competition. The threshold used for deleterious (in red) or beneficial (in blue) effect is at least 1% (−4.5 > z score > 4.5). (C) Selection coefficients for the validation of the 180 genes with significant effects measured by flow cytometry (Dataset S2). The labels are for genes with the strongest deleterious and beneficial effects. The bars indicate the SD of three biological replicates. The black circles highlight genes with haploinsufficient phenotypes. Spearman’s correlation coefficient is indicated at the Top. (D) A comparison of fitness effects among haploinsufficient and haplosufficient genes (12). P value from a Fisher exact test is shown. The fraction and number of genes are indicated with white numbers. (E) Selection coefficients of genes that code for proteins that are members of complexes and proteins that are not. On Top, we show the P value from a Wilcoxon rank-sum test.

We chose a conservative threshold of at least 1% of fitness effect (−0.01 > s > 0.01), that corresponds to a |z score| > 4.5. For most genes (86%), duplication has little or no effect on relative fitness. However, around 9% (Fig. 1B) of the duplications have moderate to strong deleterious effects (s < −0.01, z score < −4.5), while 4% have beneficial (s > 0.01, z score > 4.5) but often modest effects. We validated the top 180 genes (top 180 absolute z scores) using the same growth conditions but monitoring the two populations by flow cytometry. We generated the strains de novo and tested three biological replicates per sample. While less scalable, this approach allows for a more accurate estimation of population ratios since fluorescence is measured at the single-cell level instead of the population level. There is a strong correlation between the two assays (Spearman’s r = 0.83, P < 1e-15; Fig. 1C), giving us good confidence in the accuracy of our large-scale measurements. We validated most of the effects of the first competition assay (159/180; Fig. 1C). We compared our results with two other studies evaluating the relative fitness of strains harboring the same plasmids through pooled assays (7, 25). We found a weak but significant correlation between the selection coefficient (s) from these studies and ours (SI Appendix, Fig. S1). Although these two previous studies have used a similar pooling approach, they are only weakly correlated with each other (r = 0.15, P < 1e-7; SI Appendix, Fig. S1B). The weak correlations could come from the different media used (SC[MSG]-ura versus rich media, or nutrient limitation). Furthermore, pool assays create a complex and distinct environment compared with pairwise competition assays and may lead to noisier estimates, for instance, for strains that are rare in the pools.

The 4% of the duplications that have beneficial effects above 1% include genes such as CDC25 and IRA1 (Fig. 1 B and C), which regulate the Ras-cAMP pathway (26, 27). In yeasts, growth and metabolism in response to nutrients, particularly glucose, are regulated to a large degree by this pathway. Since we used glucose as a carbon source, duplication of these two genes may modify the activity of the pathway in a way that increases growth rate. Adaptation in limited glucose conditions often involves mutations in these pathways (28). The duplication of some genes that encode central metabolism enzymes were also beneficial. For instance, duplication of HIP1, a histidine transporter (29), and PDC2, a pyruvate decarboxylase (30), may result in increased growth rate through metabolic activity in a similar manner to duplication of genes in the Ras-cAMP pathway.

The presence of beneficial duplications may appear surprising since the yeast lineage has a high rate of duplication (31) and has undergone a whole-genome duplication (WGD) (32), which would have provided the mutational input needed to fix any beneficial duplication. One possible reason for the lack of duplication of the genes we detect with beneficial effects is that the adaptive value of some gene duplications is highly dependent on environmental conditions and can even become deleterious in specific contexts (33). Such antagonistic pleiotropy has been observed in the study of aneuploidies (34) and of gene deletions (35). We, therefore, performed a parallel competition experiment in a condition of salt stress. We find a significant correlation in the selection coefficient between conditions (Spearman’s r = 0.48, P < 1e-15), and a similar number of deleterious (8.5%) and beneficial (4.6%) effects (SI Appendix, Fig. S2). Many of the effects may, therefore, be general, while others are condition specific. The overlap in the identity of the deleterious genes between conditions is greater than for advantageous ones, suggesting that beneficial effects are more condition specific than deleterious ones. To further validate the condition specificity of the benefit of gene duplication, we measured the relative fitness of five strongly beneficial and three strongly deleterious duplications in five different conditions. We observed antagonistic pleiotropy for some beneficial duplications (SI Appendix, Fig. S3). For instance, CDC25 and IRA1 were beneficial in the standard condition, osmotic stress and with galactose as carbon source, but are strongly deleterious in 6% of ethanol, while having no detectable effect in the presence of caffeine. On the other hand, ACT1 and TUB2 have similar deleteriousness across the five conditions tested when duplicated. The mechanisms by which an increase in gene dosage is sometimes beneficial remains to be determined. In the presence of antagonistic pleiotropy, it is possible that expression in any given condition is not optimal but rather represents a tradeoff in terms of adaptation across conditions. Indeed, a study looking at the fitness effects of changes in expression levels of several genes showed that the WT expression level in some conditions is often not optimal (36).

Deleterious duplications are more frequent and have stronger effects than advantageous ones and include genes such as TUB1, TUB2, and ACT1. The products of these genes are involved in the structural integrity of the cell cytoskeleton and have been previously shown to be highly sensitive to dosage increase (37⇓–39). Our results confirm recent observations that doubling or halving the expression of TUB1 and TUB2 is enough to reduce fitness (36). These observations suggest that genes that are deleterious upon duplication could also be haploinsufficient. We indeed find that 19% of haploinsufficient genes identified by Deutschbauer et al. (12) and tested here produce deleterious effects when duplicated (17/88; Fig. 1D), which is more than the 8% expected by chance (61/811; Pearson’s χ² test, P < 1e-3; Fig. 1D). Conversely, Qian et al. (4) suggested that haploinsufficient genes may be beneficial when duplicated. We see a tendency in this direction, but it is not significant (6/88 and 34/811; χ² test, P = 0.388; Fig. 1D).

Our interpretation regarding the fitness effects of duplication depends on whether gene expression from a centromeric plasmid approximates the dosage effect of gene duplication, which we expect to commonly be a doubling of dosage [exceptions have been shown, with higher expression than expected (40)]. Centromeric plasmids usually segregate as yeast chromosomes and are on average found in one copy per haploid cell (41), but it is possible that the number of plasmids varies (42). Payen et al. (7) confirmed that in glucose-limiting conditions, the plasmids we used are typically found in only one copy per cell, so our results are overall representative of individual gene duplication events. We also examined if the plasmid-copy genes are regulated similarly as the genome-encoded copy. We reasoned that if pCEN are systematically present in multiple copies, a protein expressed from a plasmid would result in higher expression than when expressed from the genome in an equivalent genetic background where we have a genomic copy and a plasmid copy. We compared protein abundance of a strain with an endogenous green fluorescent protein (GFP) fusion and a copy of the gene on a pCEN plasmid, with the same strain but this time having the GFP fusion expressed from the pCEN plasmid. Only one gene out of five showed higher protein abundance when expressed on the plasmid, and one showed reduced expression (SI Appendix, Fig. S4A). For the two genes with different expression levels, the differences are rather modest (less than onefold) as opposed to orders in magnitude that are common in multicopy plasmids. We also directly measured the abundance of Act1p in the presence of an additional gene copy on a plasmid (pCEN-ACT1) by Western blot and found only a modest increase in abundance (SI Appendix, Fig. S4B), which is inconsistent with its deleterious effects being driven by severalfold expression changes. These observations suggest that our systematic strategy using pCEN constructions is a good experimental approximation of naturally occurring duplications in terms of dosage.

The Duplication of Individual Subunits Rarely Affects PPIs in Complexes.

The dosage balance hypothesis predicts that both underexpression and overexpression of a protein complex subunit would cause deleterious phenotypes because dosage perturbation affects the stoichiometric balance between the subunits of the complex and compromises its assembly (9, 10). For instance, some subunits of the RNA polymerase II (Rpb2p) and of the proteasome (Rpn3p) are both haploinsufficient and deleterious when duplicated. This indicates that such proteins are sensitive to changes in gene dosage in both directions, just as the gene balance hypothesis predicts. However, when we mapped the fitness effects of gene duplication on annotated yeast protein complexes, we found no significant difference compared to genes that are not in complexes (Wilcoxon rank-sum test, P = 0.28; Fig. 1E). In fact, both gene categories share a similar percentage of genes with deleterious effects (8% and 9%, respectively). These findings are robust to different z-score thresholds (Dataset S3). Therefore, the duplication of members of protein complexes is not particularly associated with a decrease in fitness, in apparent contradiction with the dosage balance hypothesis.

Our results suggest that either doubling gene dosage does not affect the assembly of protein complexes, or it affects their assembly but without having particularly strong effects on fitness. Therefore, we next aimed at directly measuring if gene duplication affects PPIs within complexes in vivo. We selected complexes that are sensitive to some but not all gene dosage changes (the 26S proteasome and the three RNA polymerases; SI Appendix, Fig. S5). We measured pairwise PPIs between all pairs of subunits with and without the duplication of each subunit using a protein-fragment complementation assay (PCA) based on the dihydrofolate reductase (DHFR) enzyme (DHFR-PCA) (43). The quantitative nature of PCA allows us to estimate a perturbation score (ps) as a direct measure of the effects of gene duplication on the PPI network of a complex (Fig. 2A).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Most duplications of subunits do not affect PPIs in large complexes. (A) DHFR-PCA–based strategy to measure perturbation of the pairwise physical interactions after a duplication. On the Left, in a DHFR-PCA, the colony size on selective media (MTX) is correlated with the stability and strength of the physical interaction between the two subunits S1 and S2 (shown in green). The perturbation score (ps) is defined as the colony size difference between the strain carrying a duplication (+pCEN-VPS5) and a WT (+pCEN) strain. Heatmap indicating ps values of the complete retromer PPI network due to the duplication of VPS5. (B) Distributions of ps for interactions with and without a competing subunit. Since the duplicated protein is not tagged with a DHFR fragment, it titrates PPI partners away from the tagged copy, decreasing colony size. We show violin plots of the distributions of all the interactions tested for five complexes (Dataset S5). On Top, we show the P value from a Wilcoxon rank-sum test. (C) Colony sizes of all strains carrying the duplication of a subunit compared with their control strain (the empty vector). Colony sizes of diploid strains carrying all tested combinations of preys, baits, and duplications for the proteasome, and the three RNA polymerases (Datasets S4 and S5). The dark gray circles indicate strains above our growth threshold indicative of physical interaction, while the light gray circles are strains below the growth threshold. The black circles indicate interactions with a competing subunit above the threshold. (D) Cumulative frequency of ps of the proteasome and the three RNA polymerases. All competition effects were excluded. Labeled are the prey–bait combinations that are perturbed by the duplication of PRE7. (E) Relationship between the selection coefficient and the average ps (absolute value) of duplicated subunits on PPIs. Only significant (FDR of 5%) and noncompetition combinations were used to calculate the averages. The circles represent duplications of proteasome subunits, while triangles represent subunits of any of the three RNA polymerases. In red, we show Spearman’s correlation coefficient.

Although nonessential retromer genes were not included in the fitness assays, we first tested our approach on this small and well-characterized complex as a proof of concept, since it has been reported to have PPIs sensitive to gene deletion (44). This complex is associated with endosomes and is required for endosome‐to‐Golgi retrieval of receptors (e.g., the Vps10 protein) that mediate delivery of hydrolases to the vacuole (45). Functionally, the retromer is divided into two subcomplexes: a cargo‐selective trimer of Vps35p, Vps29p, and Vps26p, and a membrane‐bending dimer of Vps5p and Vps17p. DHFR-PCA detects significant interactions from all the retromer subunits (SI Appendix, Fig. S6A). From our data, we detected same-subunit competition effects between DHFR-fused and nonfused copies of the proteins. Since the extra copy of the duplicated subunit is not tagged with a DHFR reporter fragment, it competes with the tagged copies for the same partners, resulting in a decreased colony size for all the interactions of this subunit (Fig. 2A). For instance, we see a reduction of the interaction of all Vps5p PPIs upon duplication of VPS5 (blue row and blue column, Fig. 2A). We also find that the duplication increases the strength of the interactions between Vps35p-Vps26p, members of the other subcomplex, the cargo-selective trimer (Fig. 2A). These results show that our strategy has enough resolution to detect small perturbations in the PPIs after the duplication of a single gene coding for a complex subunit.

We next examined the effect of gene duplication on four complexes with proteins for which some gene duplications alter fitness, as identified in our previous analysis, namely the proteasome and the three RNA polymerases (SI Appendix, Fig. S5). The proteasome is a highly conserved and thoroughly described eukaryotic protein complex that is amenable to study by DHFR-PCA (43, 46). In yeast, the core complex (20S) is associated with the regulatory particle (19S) to form a large complex (26S) composed of 37 subunits (47). From hereinafter, we will refer to the 26S proteasome as the proteasome. We tested pairwise interactions among 21 subunits as baits and 16 subunits as preys that belong to either the regulatory particle or the core complex for a total of 305 combinations. We detected 47 PPIs between subunits in the WT strain (SI Appendix, Fig. S6B). The RNA polymerases are also well-described large complexes: RNApol I includes 14 subunits, RNApol II has 16, and RNApol III has 18 subunits (48⇓–50). We tested all combinations between all available subunits since five subunits are shared between the three RNA polymerases. We observed 33 significant PPIs out of 689 combinations tested between 31 baits and 26 preys in a WT background (SI Appendix, Fig. S6C).

We observed same-subunit competition effects (Fig. 2B; Wilcoxon’s rank-sum test, P < 2e-16), which validates that additional copies of the proteins are expressed and that we can measure quantitative changes in their PPIs. Indeed, 135/181 of cases where the duplicated subunit is involved in the PPI tested show a reduced interaction score. We observed that most subunit duplications have small to nondetectable effects on the interaction network of their complex and are weakly correlated with a fitness effect. For the proteasome, excluding competition combinations, only 46 out of 8,917 combinations tested were significantly different (false discovery rate [FDR] of 5%) from the WT interactions within the complex. For the three RNA polymerases, only 28 out of 21,341 combinations tested were significantly different (Fig. 2C). Overall, most of the significant perturbations are gains of PPIs (55/74), which may suggest that when a duplication alters the interaction dynamics within the complex it does so by increasing the strength or amount of PPIs of other subunits (Fig. 2 C and D). The strongest effects are seen for the duplication of PRE7, especially for interactions Pup1p–Pre5p and Pre8p–Rpn8p (Fig. 2D). Pup1p, Pre5p, and Pre8p are part of the same subcomplex that includes Pre7p and they interact closely during the formation of the 20S proteasome. Pup1p is the β2 subunit while Pre5p and Pre8p are the ⍺6 and ⍺2 subunits, respectively (51), and share close spatial proximity, ranging from 66 to 178 Å (46).

If changes in PPIs are associated with the fitness defects we measure, we hypothesized that we would see a correlation between the perturbation of PPIs and fitness effects. We calculated the mean ps for each subunit (mean of the absolute values of significant perturbation scores) and compared it with the selection coefficient of strains containing a gene duplication of the same subunit (Fig. 2E). The correlation between ps and fitness is negative as predicted but not significant (Spearman’s r = −0.16, P = 0.49). These results suggest that subunit duplications typically have little or no effect on the protein interaction network within the complex and these effects are largely independent of the fitness effects.

Most Proteasome Subunits Have an Attenuated Expression Level When Duplicated.

Our experiments indicate that most PPIs within the proteasome interaction network and RNA polymerases are not significantly perturbed after duplication of their subunits. This suggests that these protein complexes are largely resilient to changes in gene dosage of their components. Because one of the strongest ps was observed in the proteasome, we focused on this complex to explore the underlying mechanisms of such robustness. It has been reported in multiple studies that transcription is usually correlated with gene copy number while protein abundance correlates more poorly (18). Therefore, regulatory mechanisms reducing the protein abundance of the proteasome subunits, also known as attenuation, could explain why duplication is not perturbing PPI between the subunits.

To test whether the proteasome subunits are attenuated, we looked for changes in protein abundance after gene duplication. We compared protein abundance in GFP-tagged strains (52) carrying a duplication of the gene or an empty vector as a control (Fig. 3A). Protein attenuation would lead to a reduction of protein levels of both copies, which would result in reduced fluorescence signal. Most subunits (17/19) have a significant decrease in GFP-fluorescent signal after duplication (Fig. 3B and Dataset S6). Next, we calculated an attenuation score: the difference between WT and the duplicated GFP-fluorescent signals normalized by the WT (Fig. 3C). If the expression of the fused copy was reduced by half to balance the additional copy in a plasmid (complete attenuation), the attenuation score would be 0.5. Interestingly, attenuation is similar for subunits belonging to the same subcomplex (Fig. 3D), suggesting a regulation that depends on complex assembly. Havugimana et al. (53) reported that the stoichiometry within each proteasome subcomplex is 1:1, while the stoichiometry among subcomplexes varies from 1:1 to 1:4 (Dataset S6). This observation could explain why subunits belonging to the same component have similar expression and attenuation patterns, while there are significant differences between components (Fig. 3D; P = 0.002, one-way ANOVA test).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

Attenuation of protein abundance after duplication in most proteasome subunits. (A) Measure of attenuation with GFP-tagged proteins. Changes in abundance of each subunit can be detected by comparing fluorescent signals of GFP-tagged subunits before and after duplication. Upon attenuation, the abundance of the tagged copy will be reduced. (B) GFP signal comparison between strains carrying a duplication of the GFP-tagged subunit and their corresponding control. On the Right, a cartoon of the proteasome with its components. All GFP values are corrected for autofluorescence by subtracting the signal of the parental strain not expressing GFP and by cell size (Dataset S6). (C) Attenuation scores of all assayed proteasome subunits. The attenuation score is the difference between GFP fluorescent signals of the control strain (bearing a control plasmid) and the duplicated strain (bearing a centromeric plasmid with an extra copy of the subunit) divided by the GFP signal of the control. In the absence of attenuation, this value is 0. Upon complete attenuation, it is 0.5. (D) Attenuation scores of the proteasome subcomplexes. On the Right, the asterisks indicate significant differences between components calculated by correcting for multiple testing (Tukey’s test; *P ≤ 0.05 and **P ≤ 0.01). (E) Colony sizes of all strains carrying the PRE7 duplication (+pCEN-PRE7) compared with their control strain (the empty vector) indicating changes in PPI in the DHFR-PCA assay. The black dots highlight the subunits that have interactions disturbed after PRE7 duplication. (F) GFP signal of proteasome subunits before and after PRE7 duplication (+pCEN-PRE7). The black dots highlight the subunits that have interactions disturbed after PRE7 duplication (Dataset S7). Replicate measurements are available in SI Appendix.

Strikingly, all but two complex subunits appear to be attenuated. One of the most attenuated subunits Pre7p has a significant effect on both PPIs and fitness when duplicated (Fig. 2C). The duplication of PRE7 perturbs PPIs between Pre5p, Pre8p, Pup1p, and Rpn8p (Fig. 3E). All of them share close spatial proximity with Pre7p in the proteasome core complex. We tested whether PRE7 duplication affects the abundance of these subunits by measuring the GFP signal of all proteasome subunits with and without PRE7 duplication. Most subunits are unaffected upon PRE7 duplication but the four “perturbed” subunits (Pre5p, Pre8p, Pup1p, and Rpn8p) show a modest increase in their protein abundance (Fig. 3F). Even though the difference between the control and the PRE7 duplicated background is small, it is highly reproducible and significant (SI Appendix, Fig. S7A), and appears specific to the subunits with altered interactions (SI Appendix, Fig. S7B). These modest but significant changes in protein abundance of perturbed subunits after the duplication of PRE7 may explain the changes we observed in their PPIs. The duplication of PRE7, even if largely attenuated, affects the organization of the proteasome by affecting the abundance and interactions of a few other subunits. The dosage balance hypothesis may therefore apply to a limited number of subunits.

The proteasome subunits show different levels of attenuation, and recent studies (18, 19) suggest that attenuation occurs mostly at the posttranscriptional level across complexes. To examine this specifically for the proteasome, we retrieved data from Dephoure et al. (18) and looked at the mRNA abundance ratio of individual genes in disomic strains relative to WT. Most proteasome subunits roughly double their transcript levels when duplicated (relative to a log2 mRNA ratio around 1; SI Appendix, Fig. S8A). This includes Pre3p and Pre7p, two of the most attenuated subunits in our experiments (Fig. 3C). Since aneuploidies can cause systemic-level changes on the transcriptome and proteome and confound the effects of an individual duplication, we performed RT-qPCR analysis of four attenuated subunits (PRE7, RPT2, RPT6, and SCL1), including three not measured in Dephoure, a nonattenuated subunit (PRE10), and a gene that is not a part of the proteasome (FAS2). As expected, the nonattenuated genes PRE10 and FAS2 present no significant change in their transcript levels after their duplication (t test, P = 0.6 and 0.78; SI Appendix, Fig. S8B). Highly attenuated genes at the protein level exhibit varying responses: RPT6 shows a significant mRNA attenuation (t test, P = 0.01), and RPT2 and SCL1 also show a marginally significant mRNA attenuation (t test, P = 0.1 and 0.07). PRE7 displays no significant change in transcript levels (t test, P = 0.24). These results from individual gene duplication are therefore consistent with the data observed for disomic strains. Combined, these data suggest that attenuation can occur at the transcriptional level but more frequently at the posttranscriptional level (SI Appendix, Fig. S8C). In most cases, as for PRE7, attenuation appears to be posttranslational because transcription and translation rates are both maintained in disomic strains (SI Appendix, Fig. S8A) (17). We further explored the attenuation of the PRE7 subunit by using a tunable expression system. We constructed a strain with tunable PRE7 expression (54) and which contains two gene copies of PRE7 tagged with different fluorescent proteins that were independently monitored by flow cytometry. There is a significant negative correlation (Pearson’s correlation r = −0.25, P < 10e-15) between the protein abundance of the copy expressed under the inducible promoter and abundance of the chromosome copy (SI Appendix, Fig. S9 and Dataset S8). These results suggest that attenuation is expression level dependent. Finally, we examined whether this posttranslational control depends simply on the presence of an additional gene copy or on that gene copy being transcribed and translated. For this purpose, we generated a centromeric construction that contains PRE7’s full promoter and open reading frame (ORF) but that cannot be translated because it lacks a start codon (two constructions) or the start codon is followed by a stop codon (SI Appendix, Fig. S10). Pre7p remains at the WT level of expression in the presence of these three constructions. Since all the cis-regulatory elements remain intact on these constructions, it is fair to assume that transcription is initiating but translation is not initiated or is terminated prematurely due to the insertion of the stop codons. This confirms that attenuation requires the translation of the mRNA.