A gene expression library reveals costs of single gene duplication

Previous studies strongly suggested that aneuploidy toxicity is dependent on genes encoded on the amplified chromosomes, since large DNA segments without coding potential are not especially toxic in yeast [4,33]. But whether aneuploidy toxicity, and Ssd1 dependence in particular, is due to only a few specific genes or the cumulative burden of many genes was unclear. To distinguish between models and study Ssd1’s role in aneuploidy tolerance, we measured the cost of individual gene duplications, using a barcoded plasmid library (Fig 2A). The MoBY 1.0 gene library comprises ~5,000 open reading frames along with their native upstream and downstream sequences, cloned onto a barcoded centromeric (CEN) plasmid [40]. We transformed the library into wild-type and ssd1Δ euploid strains as well as select wild-type and ssd1Δ aneuploids with duplications of Chr4, Chr7, Chr12, and Chr15 (the Chr15 wild-type aneuploid was omitted due to poor library coverage). These strains were chosen because they present a range of growth defects and varying degrees of Ssd1 dependence. Each copy of the duplicated chromosomes carried a different selectable marker, allowing both copies of the chromosome to be maintained throughout the experiment. Each pooled library was grown for 10 generations (with the exception of the YPS1009 Chr4 ssd1Δ strain, which was grown for only 5 generations due to its extreme fitness defect), in at least 3 biological replicates (see Methods). An aliquot of the library was collected before and after competitive growth, and the relative abundance of each plasmid in each sample was scored by sequencing the plasmid barcodes (see Methods). Barcodes that decreased in abundance during competitive growth correspond to deleterious gene duplicates, whereas barcodes that increased in abundance represent genes whose duplication is advantageous. To quantify fitness effects, we calculated a fitness score for each gene duplication, defined as the log₂(fold change) of normalized barcode counts after competitive growth (see Methods). We applied several types of comparisons to understand the impact of aneuploidy in SSD1+ and ssd1Δ cells (Fig 2B). Data for the euploid YPS1009 were first reported in a parallel paper from our lab [32].

Open in a separate window

Fig 2

Distribution of fitness costs of gene duplication are similar in aneuploids and euploids.

A) Schematic of the approach, see text for details. B) Library competition experiments were performed in strains indicated in the diagram. Specific comparisons (indicated by a number) were made to compare: (1) Wild-type aneuploids versus wild-type euploid; (2) ssd1Δ euploid versus wild-type euploid; (3) ssd1Δ euploid versus wild-type aneuploids, used for modeling; (4) ssd1Δ aneuploids versus wild-type euploid, ssd1Δ euploid, and corresponding wild-type aneuploid when available. C) Distribution of fitness scores (log₂(fold change) of normalized barcode counts) for all measured genes in euploid (Eu.) and aneuploids with the denoted chromosome amplification. D) Distribution of pairwise differences in log₂ fitness scores between each wild-type aneuploid and batch-paired euploids (Comparison 1 in B), for 2,403 detrimental genes identified in the euploid wild-type [32]. Eu_B2-B1 represents a control distribution comparing fitness scores in two batches (B1 and B2) of euploid replicates. https://commons.wikimedia.org/wiki/File:Erlenmeyer_Flask_%2822792%29_-_The_Noun_Project.svg.

As reported in this parallel paper, duplication of 2,403 genes produced a significant fitness effect in the euploid strain at a false discovery rate (FDR) < 0.05 [32]. Re-analysis here showed that beneficial gene duplicates were enriched for essential genes (p = 9.1E-08, hypergeometric test). Genes whose duplication was detrimental were not enriched for any gene ontology terms; however, we previously noted that deleterious genes tended to be longer and encode proteins with more disorder [32], consistent with other reports [41]. There was no enrichment for proteins in multi-subunit complexes and no difference in the number of protein interactions for deleterious genes compared to neutral genes. None of the chromosomes showed a significant overrepresentation of detrimental genes that passed FDR significance (FDR > 0.05, hypergeometric test with Benjamini and Hochberg correction).

Aneuploid strains are not universally more sensitive to gene duplications

Here we focus on the effects of single-gene duplication in aneuploid strains. We initially hypothesized that aneuploids may be more sensitive to the gene duplication library, since they already carry a burden of extra gene content. If aneuploids were more sensitive to gene duplication, we would expect more deleterious genes in aneuploid strains and/or more severe fitness defects upon duplication. On a global scale, this was not the case: the proportion of detrimental genes was comparable between aneuploids and batch-paired euploids (see S1 Table), albeit with a marginally higher proportion of detrimental genes in aneuploids as seen in comparing the distributions (Fig 2C). This could not be explained by data normalization issues, since we confirmed that genes scored as neutral in the library experiment had no significant effect on growth rate when strains were grown in isolation (S1 Fig). Although there was a longer tail of genes with stronger negative effects in the aneuploid strains (Fig 2D), the bulk of the distribution was similar to euploids. We conclude that the aneuploid strains are not sensitized to over-expression of all genes, but they are overly sensitive to a subset of duplicated genes.

To investigate further, we identified 335 genes that were deleterious in one or more aneuploids (FDR < 0.05) with a fitness score at least 1.5 in log₂ space more deleterious than batch-paired euploid measurements (Comparison 1 in Fig 2B, see Methods and S1 Data). Of these 335 deleterious gene duplicates, 58 were common to two or more aneuploid strains, suggesting common effects across aneuploid strains. There was no enrichment for any gene ontology terms. However, we noticed that 8 of the common detrimental genes are naturally encoded adjacent to a centromere, such that the centromere was cloned onto the plasmid along with the gene. Dicentric plasmids are known to be unstable in yeast [42], likely because spindle attachment to both centromeres results in DNA breakage [43]. The number of especially deleterious plasmids that carried a cloned centromere was higher than expected by chance (p = 5.5E-09, hypergeometric test). This raised the possibility that plasmids with multiple centromeres may be especially unstable in aneuploid strains relative to euploids.

In contrast, we noticed that some genes scored as deleterious in the euploid were much less deleterious in aneuploids (Fig 2D). We identified 360 genes that were deleterious in the euploid strain (FDR < 0.05) but at least 1.5 in log₂ space less so in one or more wild-type aneuploids (Comparison 1 in Fig 2B and S1 Data). 153 of these genes were common to two or more aneuploid strains, again suggesting common effects independent of which chromosome is duplicated. Intriguingly, this group was enriched for genes involved in the mitotic spindle (p = 9.1E-05, hypergeometric test). Previous work by Storchova et al. [44] showed that a subset of genes affecting the mitotic spindle are essential in tetraploid, but not diploid, yeast. The authors proposed that increased chromosome content may alter requirements for spindle geometry for proper chromosome management. Our results hint that aneuploid cells with extra chromosomes may benefit from increased abundance of spindle components. In sum, our results suggest that aneuploid strains are not generally more sensitive to further gene duplication but that a subset of genes are more or less toxic in aneuploids.

Specific gene duplications are especially toxic in ssd1Δ euploids and explain a portion of Ssd1 aneuploidy dependence

As shown above, SSD1 deletion sensitizes cells to aneuploidy to varying degrees depending on which chromosome is amplified. Given Ssd1’s role in translational suppression, we wondered if SSD1 deletion sensitizes cells to gene over-production in general. We therefore used the library to measure fitness costs of single-gene duplication in the ssd1Δ euploid. Overall, the distribution of fitness scores for gene duplications was not significantly different than the euploid wild type (p = 0.38, Wilcoxon rank-sum test, Fig 3A), indicating that ssd1Δ does not sensitize cells to all gene duplications. We directly compared ssd1Δ euploid versus wild-type euploid cells (Comparison 2 in Fig 2B) and identified 218 genes more detrimental in the ssd1Δ euploid (FDR < 0.05, Fig 3B and S1 Data). We were especially interested in identifying functional or biophysical properties of these genes; however, the group did not have striking functional enrichments, aside from 4 genes involved in ER to Golgi transport (p = 1.7E-03, hypergeometric test). We tested more than a dozen other transcript or protein features including abundance, structure or sequence composition, or other biophysical properties, but found nothing significant (see Methods). Furthermore, the group was not enriched for mRNAs bound by Ssd1 (p = 0.67, hypergeometric test) or that harbor the known Ssd1 binding motifs [22,24]. Reciprocally, genes encoding Ssd1-bound mRNAs were not more deleterious when duplicated in the absence of SSD1 (Fig 3C). The only significant feature we found was that the length distribution of these 218 genes is shorter than all other genes in the dataset (p = 1.3E-04, Mann-Whitney U test). This is notable since genes whose duplication is deleterious in the wild type tend to be longer than neutral genes [32]. Thus, it is unclear why these 218 gene duplicates are more deleterious in ssd1Δ cells.

Open in a separate window

Fig 3

A subset of genes are more detrimental in ssd1Δ euploid cells relative to wild type.

A) Distributions of fitness scores for all 4,462 measured genes in wild-type euploid and ssd1Δ euploid strains. B) Scatterplot of replicate-averaged fitness scores for all measured genes in ssd1Δ euploid (y-axis) and wild-type euploid (x-axis); 218 genes significantly more deleterious in ssd1Δ euploid cells are shown in pink. C) Fitness scores for 336 genes encoding mRNAs bound by Ssd1 (see Methods) in wild-type (blue) and ssd1Δ (light blue) euploid strains. D) Ssd1 dependence as shown in Fig 1 (y-axis) plotted against the predicted Ssd1 dependence, calculated as the sum of log₂ gene-specific Ssd1 dependence values as described in the text (x-axis), over the subset of 218 genes whose duplication is toxic to ssd1Δ cells. E) Ssd1-dependent fitness scores for the 218 genes analyzed in (D), plotted according to encoding chromosome.

Nonetheless, we used these fitness scores to further understand models of Ssd1 dependence. In other recent work from our lab, we showed that nearly 70% of the variance in wild-type aneuploid growth rates can be explained by the cumulative fitness costs of genes duplicated on each chromosome [32]. Here we used a similar approach to explain Ssd1 dependence. We tested if the Ssd1 dependence for each chromosome duplication (as shown in Fig 1) could be explained by the cumulative Ssd1 dependencies of genes duplicated on that chromosome. We calculated Ssd1 dependence for each gene as the fitness cost for that gene duplication in the ssd1Δ strain versus the fitness cost for that gene duplication in the SSD1+ euploid strain. We then calculated the per-chromosome Ssd1 dependence as the sum of relative log₂ fitness scores over the 218 genes more deleterious in the mutant strain (see Methods). This simple model explained roughly half of the variance in Ssd1 aneuploidy dependence, with an adjusted R² of 0.46 (Fig 3D, p-value = 3.2E-03). Omitting Chr16 from consideration increased the explanatory power to 51% (adj. R² 0.51). Model performance decreased when only the top 10% of ssd1Δ -sensitive genes were used (36% of variance explained, S2 Fig). Thus, Ssd1 dependence for tolerating chromosome duplications can be partly explained by the cumulative burden of specific genes on each chromosome; however, this number is more than one or two genes per chromosome (Fig 3E).

Many gene duplicates toxic in ssd1Δ euploids are also toxic to SSD1+ aneuploids but not SSD1+ euploids

Ssd1 is required to tolerate specific gene duplications, although the reasons are unclear. We wondered if SSD1+ aneuploids may be sensitive to the same set of genes. We explored this in several ways. First, we realized that genes more deleterious in one or more aneuploid strains (Comparison 1 in Fig 2B) are enriched for the 218 genes whose duplication is especially deleterious in ssd1Δ euploid cells (Comparison 2 in Fig 2B, p = 2.9e-06, hypergeometric test). Second, we interrogated the fitness scores of these 218 genes in wild-type aneuploid cells. We plotted the average fitness scores from the three wild type aneuploids against the euploid fitness scores for these 218 genes (Fig 4A). Indeed, 145 (67%) of these genes were more deleterious in aneuploids than in the euploid. One subset of genes was especially more toxic in aneuploids, indicated by their deviation from the linear trend (Fig 4A, labeled points). We confirmed that these genes were more deleterious in individual aneuploids compared to the euploid (Fig 4B). Thus, many of the gene duplications that are toxic in ssd1Δ euploids are also more deleterious in SSD1+ aneuploids versus the euploid. This indicates that aneuploidy may be mimicking the cellular state of ssd1Δ euploid cells (see Discussion).

Open in a separate window

Fig 4

Many gene duplications toxic in ssd1Δ euploids are also toxic in SSD1+ aneuploids.

A) Average wild-type aneuploid fitness scores for each gene (y-axis) plotted against the corresponding fitness score measured in wild type euploid cells (x-axis), for the 218 genes in Fig 3B. Genes with especially deleterious effects in aneuploids relative to the euploid are labeled with gene name. B) Fitness scores for genes labeled in Fig 4A (rows) in biological replicates of each strain (columns), colored according to the key. C) The relative fitness of aneuploid strains (as in Fig 1) was plotted against the predicted relative fitness, taken as the sum of log₂ fitness values for the 218 gene duplicates as measured in the ssd1Δ euploid strain (x-axis).

We wondered how well the cumulative cost of these 218 gene duplicates, identified by their sensitivity in the ssd1Δ euploid, could explain chromosome fitness costs in SSD1+ aneuploids. We calculated the fitness cost of each chromosome duplication in SSD1+ cells as the cumulative cost of these 218 gene duplicates, based on measurements from the wild-type euploid strain (see Methods). The calculated fitness costs did not agree with the measured costs in the SSD1+ background (adjusted R² = 0.07, p-value = 0.18). However, many of these genes are only deleterious when duplicated in the ssd1Δ euploid strain. This raised the possibility that the fitness costs of these gene duplicates in the ssd1Δ euploid was most relevant. We therefore modeled the fitness cost of chromosome duplication in SSD1+ cells, but based on gene fitness costs measured in the ssd1Δ euploid strain. Remarkably, this model explained 73% of the variance in growth rates of the SSD1+ aneuploids (Fig 4C). This important result indicates that chromosome duplication in an SSD1+ background may mimic the liabilities of SSD1 deletion in a euploid strain, sensitizing cells to the same subset of problematic gene duplications (see Discussion).

Duplicated genes beneficial in ssd1Δ aneuploids implicate mechanisms of aneuploidy tolerance

Since SSD1 deletion sensitizes strains to the stress of chromosome duplication, we wondered if genes beneficial in ssd1Δ aneuploids might complement processes that are defective in those aneuploids. We therefore identified genes whose duplication compensates growth defects in ssd1Δ aneuploids under the conditions used here (Comparison 4 in Figs ​Figs2B2B and ​5A5A). There were 445 genes whose duplication was beneficial to at least one ssd1Δ aneuploid strain (FDR < 0.05) and that displayed a higher fitness score than observed in corresponding wild-type aneuploids and both euploids (see Methods and S1 Data). We identified the greatest number of beneficial genes in ssd1Δ cells with an extra copy of Chr12, possibly due to increased statistical power from 4 replicates (but perhaps also influenced by the rDNA locus on this chromosome [83]), followed by ssd1Δ Chr4 aneuploid cells. The fewest beneficial duplications were found in the ssd1Δ Chr7 aneuploid, which is the least dependent on Ssd1 of the aneuploids studied here. To validate, we tested 13 plasmids expressed in Chr12 aneuploids; 8 of the 13 genes significantly improved growth in the ssd1Δ aneuploid relative to empty vector, confirming library measurements (S3A Fig).

Open in a separate window

Fig 5

Gene duplications that benefit ssd1Δ aneuploids implicate translation.

A) Hierarchically clustered fitness scores for 445 gene duplications (rows) beneficial in at least one aneuploid strain (columns), colored according to the key. Fitness scores represent the average of at least biological triplicate. (*) denotes scores measured after 5 generations instead of 10 generations, see text. B) Venn diagram of beneficial genes identified in ssd1Δ aneuploids. Dotted border identifies beneficial genes common to 3 or more ssd1Δ aneuploids. The inset lists 18 genes identified in 3 or more ssd1Δ aneuploids; bolded genes have known links to translation. C) Average and standard deviation of relative growth rates of the denoted strains grown in 1ug/mL nourseothricin (NTC) compared to a water control (n = 3). Measured growth rates are indicated as solid bars; expected growth rates based on the combined effect of NTC and underlying aneuploidy growth defect are shown as dashed bars (see Methods). (**) p-value < 0.01, (*) p < 0.05, (+) p < 0.1, one-tailed paired T-test.

Comparing across strains revealed that many gene duplications were beneficial to multiple ssd1Δ sensitized aneuploids (Fig 5B, see Methods). We were especially interested in these genes, which may point to processes that are generally compromised upon chromosome duplication. 98 of these gene duplications showed positive fitness scores in at least 2 aneuploids, and 18 duplications were positive in at least 3 aneuploids. The group of 98 was enriched for genes encoding small GTPase-mediated signaling proteins (p = 3.7E-05, hypergeometric test) and mRNA processing proteins (p = 9.8E-04), while the group of 18 was enriched for translational regulators (p = 1.1E-03, hypergeometric test). Somewhat surprisingly, we did not find enrichment of genes linked to protein folding or the ubiquitin-proteasome system, which were previously associated with aneuploidy tolerance in the ssd1- W303 strain [45]. Of the five plasmids with positive fitness scores in all four aneuploid ssd1Δ strains, four encode proteins directly implicated in translational regulation: poly-A binding protein Pab1, translational repressor Sbp1, and eIF5a elongation factor Hyp2, which was cloned along with uncharacterized gene UTR5 on two of the selected plasmids. The fifth gene, GLC7, encodes PP1 phosphatase that can act on translational regulator eIF2a in yeast [46], as well as on other proteins. We confirmed that PAB1, SBP1, and HYP2/UTR5 all significantly improve growth rates of ssd1Δ Chr12 aneuploid cells relative to empty vector (p < 0.05, S3A Fig). We tested the HYP2/UTR5 plasmid in other strains and found that it also benefited ssd1Δ with Chr4 duplication and, to a minor degree that just missed significance (p = 0.08), the ssd1Δ strain with Chr7 duplicated (S3 Fig).

We performed a similar analysis for the wild-type aneuploids, and identified 66 genes that provided a benefit to two or more wild-type aneuploids relative to batch-paired euploid (S1 Data). Interestingly, ten of these genes overlap the 98 genes that were beneficial to multiple ssd1Δ aneuploids (S4 Fig). One of these was poly-A binding protein encoded by PAB1—indeed, duplication of PAB1 provided a mild but significant benefit to the tested wild-type YPS1009_Chr12 strain (S3A Fig).

MoBY 1.0 competitive growth

Competition experiments were performed as in [40,83]. Briefly, 1 mL frozen glycerol stocks of library transformed cells were thawed into 100 mL of liquid SC-His with NTC (100 mg/L) and G418 (200 mg/L) at a starting OD₆₀₀ of 0.05, then grown in shake flasks at 30°C with shaking. An aliquot of frozen stock represented the starting pool (generation 0) for each strain. After five generations, each pooled culture was diluted to an OD₆₀₀ of 0.05 in fresh media to maintain cells in log phase. At 10 generations (5 generations where indicated) cells were harvested and cell pellets were stored at −80°C.

Data were collected in two batches. Batch 1 data were generated from two independent competitive growths from each of two independent transformations of the plasmid library, for a total of 4 replicates. Reproducibility of fitness scores was consistently high between independent growths from the same transformation stock (Pearson correlation coefficients PCC 0.87–0.96) and more modest across independent growths from different transformation stocks (PCC 0.44–0.67). Thus, to capture this variation, Batch 2 included three independent competition experiments, each from a different transformation of the pooled library. This ultimately yielded noisier fitness scores but greater confidence in reproducible effects. We pooled euploid controls from Batch 1 and Batch 2 for 7 replicates each of the wild-type euploid [32] and ssd1Δ euploid experiments. Fitness scores generated from Batch 1 and Batch 2 were reproducible, with PCC of 0.93 and 0.92 for wild-type euploid and ssd1Δ euploid, respectively.

MoBY 1.0 sequencing

Plasmids were recovered from each pool using Zymoprep Yeast Plasmid Miniprep II (Zymo Research D2004-A). Plasmid barcodes were amplified using primers described in [84] and 20 PCR cycles. Barcode amplicons were pooled and purified using AxyPrep Mag beads (1.8X volume beads per sample volume) according to the manufacturer’s instructions (Axygen). Pooled amplicons were sequenced on one lane of an Illumina HiSeq 4000 to generate single-end 50 bp reads. Reads were mapped to 4,872 unique upstream tags (barcodes) [40]. This resulted in 126.2 million reads with exact matches to MoBY 1.0 barcodes in Batch 1 across 32 samples (average of ~3.9M reads/sample) and 154.7 million reads in Batch 2 across 51 samples (average of ~3.0M reads/sample). Sequencing data and barcode counts are available in the Gene Expression Omnibus (GEO) under accession GSE263221.

MoBY 1.0 dataset analysis

To limit bias from barcodes that were under-represented in the starting pool, we removed barcodes in the bottom 5% of total counts at generation zero or with 0 reads across all samples (n = 303 for Batch 1 comparisons, n = 338 for Batch 2 comparisons, n = 313 for integrated comparisons across all 7 euploid replicates). A pseudocount of 1 was added to every measurement, and resulting data were normalized by TMM normalization [85] in edgeR version 3.36.0 [86]. TMM normalization produced very similar results to normalization by total barcodes sequenced. Barcodes that changed significantly in abundance in each experiment (i.e. after 10 generations of outgrowth compared to each starting pool) were identified using a glmQ function in edgeR with quasi-likelihood tests, using Benjamini and Hochberg multiple test correction and taking a false discovery rate (FDR) < 0.05 as significant [87]. A small number of genes cloned on multiple plasmids, and thus reported with multiple barcodes, were removed from downstream analysis. In total, barcodes representing 4,508 genes were analyzed across the two batches of experiments, interrogating indicated aneuploids and batch-paired euploids. Fitness scores correspond to the average log₂(fold change) reported in edgeR output files. Hierarchical clustering was performed using Cluster 3.0 [88] and visualized using Java TreeView [89]. 5 generation collections (Chr4 ssd1Δ, Chr4 wild-type, and isogenic euploids) were analyzed separately. To maximize available statistical power, we combined and reanalyzed all wild-type and ssd1Δ euploid replicates from Batch 1 and Batch 2 (referred to as “integrated analysis”).

Enrichments

Enrichments for GO terms were assessed using setRank version 1.0 [90] or GO Term Finder version 0.86 in the Saccharomyces Genome Database [91], comparing to a background dataset of all measured genes. Statistical significance was assessed using Mann-Whitney U tests for continuous data and hypergeometric tests for categorical terms. DNA motif searches were performed on sequences -500 bp upstream and +100 bp downstream each gene’s open reading frame, using MEME suite [92] for novel motif discovery and SEA [93] to look for enrichment of genes whose upstream regions harbor previously identified motifs associated with Ssd1 binding, AKUCAUUCCUU and CNYUCNYU, respectively [22,24]. Propensity for disorder was quantified by % residues with IUPRED3 disorder propensity score >0.5 [94]. Ssd1 targets were compiled from [21,24]; the Bayne et al. targets comprise the 189 targets identified at 30 degrees C as at least 2-fold enriched at FDR < 0.05 [24]. Other mRNA/protein features as listed in [32] were analyzed as outlined in that study, along with age-dependent ribosome pause sites from [95,96]. None of those features aside of what is listed in the main text showed statistical significance.

Modeling

Linear modeling was performed using ordinary least squares (OLS) regression as implemented in the statsmodels version 0.11.1 package for python3. Visual representations of the regressions were created using seaborn version 0.11.2, with 95% confidence interval plotted using the lmplot function. Predicted Ssd1 dependence (Figs ​(Figs11 and ​and3D)3D) or predicted aneuploid fitness relative to the euploid (Fig 4C) were taken as the sum of scores for genes indicated in each figure legend that are on each chromosome.

Thanks to Michael Place for technical assistance, Norah Paulsen and Emma Mayhew for experimental support, and to the Gasch Lab for helpful comments.